MameLoshnLM: Yiddish Language Model and Evaluation Benchmark

# Yiddish Language Model and Evaluation Benchmark
Source: [https://arxiv.org/html/2608.05850](https://arxiv.org/html/2608.05850)
Uri Katz1Omer Goldman2Tomasz Limisiewicz3 Reut Tsarfaty1,אָ\{\}^\{1,\\text\{אָ\}\}Noah A\. Smith3,4,אָ\{\}^\{3,4,\\text\{אָ\}\} 1Bar\-Ilan University2University of Cambridge 3University of Washington4Allen Institute for AI

###### Abstract

We presentMameLoshnLM,111The model and related resources are available at[https://github\.com/katzurik/MameLoshnLM](https://github.com/katzurik/MameLoshnLM)\.the first open\-source 8B\-parameter language model built specifically for Yiddish\. Despite Yiddish’s rich textual tradition, its limited digital presence and the scarcity of reliable evaluation resources have constrained progress in Yiddish language modeling\. Existing multilingual corpora and benchmarks are often poor proxies for the language, containing substantial amounts of noisy, machine\-translated, and misclassified text\. We address these gaps by introducingOytser, a high\-quality Yiddish pretraining corpus that combines contemporary web\-native sources with literary materials, andKashes, a multi\-task benchmark spanning translation, linguistic analysis, information extraction, and language understanding\. Using these resources, we continue pretraining Llama 3\.1 8B to obtainMameLoshnLM\. Across the tasks in the benchmark,MameLoshnLMoutperforms open baselines of similar scale\. Our analyses show that these gains are not only quantitative: relative to general\-purpose multilingual models,MameLoshnLMbetter captures language\-defining lexical and morphological patterns, pointing to a broader failure mode of noisy web\-scale multilingual data for low\-resource languages\. Our results provide both a foundation for Yiddish NLP and a practical template for language model development in historically rich but digitally underrepresented languages\.

## 1Introduction

Despite swift advancements in multilingual language modeling, large language models’ \(LLMs\) benefits remain highly uneven across languages\(Wu et al\.,[2025](https://arxiv.org/html/2608.05850#bib.bib37)\)\. Performance is typically strongest for languages with extensive digital presence, large quantities of high\-quality text, and mature evaluation resources, while low\-resource languages remain substantially underserved\. Yiddish is a particularly interesting case\. Historically a Germanic language written primarily in Hebrew script, it also includes a substantial Hebrew and Aramaic lexical layer and reflects long\-standing contact with Slavic languages\. Although it has a long and rich literary history, its contemporary online presence is limited, and much of the Yiddish text available in common web\-scale datasets is sparse, noisy, or poorly matched to the language as it is used today\. The result is models that serve neither the language’s roughly one million speakers nor the academic community that researches the history and culture of the Yiddish\-speaking world\.

These limitations reflect gaps across the full LM development pipeline\. On the data side, the amount of publicly available Yiddish pretraining text is limited\(Xue et al\.,[2021](https://arxiv.org/html/2608.05850#bib.bib38)\), and its quality has not been well characterized\(Kreutzer et al\.,[2022](https://arxiv.org/html/2608.05850#bib.bib21)\)\. In this paper, we show through a detailed analysis of mC4 that much of its putative Yiddish content is either machine\-translated spam or not Yiddish at all\. We find that less than half of the data is genuine Yiddish text\. On the evaluation side, benchmarks for Yiddish are scarce, and those that do exist are often based on automatic translation rather than tasks designed for the language itself\(Singh et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib31)\)\. Addressing Yiddish effectively therefore requires more than simply adding more text: it requires coordinated work on corpus construction, evaluation, model adaptation, and analysis\.

We introduceMameLoshnLM, an open\-source 8B\-parameter LM for Yiddish, together with two supporting resources:Oytser, a new Yiddish pretraining corpus, andKashes, a benchmark for evaluating Yiddish LMs\.Oytserwas designed as a higher\-quality alternative to common open multilingual resources by combining contemporary web\-native Yiddish with materials drawn from Yiddish’s literary tradition\. This is particularly important for Yiddish, whose historical textual record is far richer than its current web footprint\. We also introduceKashes, a benchmark comprising existing and newly developed tasks, intended to provide a broader and more reliable basis for evaluating Yiddish language models\.

Using these resources, we continue pretraining Llama 3\.1 8B to obtainMameLoshnLM\. Across the benchmark,MameLoshnLMoutperforms open models of similar scale, with strong gains on Yiddish\-centered tasks such as translation, linguistic analysis, and named entity recognition\. Beyond benchmark performance, our analysis shows thatMameLoshnLMoutputs more natural Yiddish, morphologically and lexically, avoiding overly Germanized and translationese\-like characteristics often produced by general\-purpose multilingual LMs \([Section 8](https://arxiv.org/html/2608.05850#S8)\)\. We further find that, while English data is important for preserving the broader capabilities of the base model during continued pretraining, adding data from historically or genealogically related languages yields only limited and inconsistent gains \([Section 7\.2](https://arxiv.org/html/2608.05850#S7.SS2)\)\.

More broadly, our results point to a challenge that extends beyond Yiddish\. For historically rich but digitally underrepresented languages, the main obstacle to language\-model development may be not only data scarcity, but mismatch between authentic language use and the noisy public web data on which multilingual models are often trained\. We present Yiddish as a clear case of this problem, andMameLoshnLM,Oytser, andKashesas a practical demonstration that targeted resource construction and continued pretraining can substantially narrow the gap\.

## 2Related Work

Recent work has shown that continued pretraining of strong open\-weight models is an effective approach for adapting LMs to low\- to moderate\-resource languages, including, for example, Basque, Estonian, Kazakh, and Setswana\(Etxaniz et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib13); Kuulmets et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib23); Koto et al\.,[2025](https://arxiv.org/html/2608.05850#bib.bib20); Brown & Marivate,[2025](https://arxiv.org/html/2608.05850#bib.bib7)\)\. However, this paradigm still leaves open important questions about training data composition\(Zhang et al\.,[2025](https://arxiv.org/html/2608.05850#bib.bib40)\), the role of language mixing\(Elhady et al\.,[2025](https://arxiv.org/html/2608.05850#bib.bib12)\), and the value of machine\-translated versus native text in low\-resource adaptation\(Doshi et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib11)\)\. Our work follows this general approach for the low\-resource language Yiddish, but examines it in a particularly constrained setting, marked by limited training data, low\-quality publicly available text, and very limited evaluation resources\.

Although Yiddish has received only limited attention in NLP, several important efforts have created the foundations on which modern work can build\. Early work addressed phrase\-based machine translation\(Genzel et al\.,[2009](https://arxiv.org/html/2608.05850#bib.bib16)\), while later efforts developed foundational infrastructure such as a Yiddish speech corpus\(Webber et al\.,[2022](https://arxiv.org/html/2608.05850#bib.bib36); Bleaman & Nove,[2025](https://arxiv.org/html/2608.05850#bib.bib6)\), and a basic ASR system\(Ćavar et al\.,[2016](https://arxiv.org/html/2608.05850#bib.bib8)\)\. Other work supported core NLP building blocks through resources and models for part\-of\-speech tagging\(Santorini,[2021](https://arxiv.org/html/2608.05850#bib.bib30); Kulick et al\.,[2022](https://arxiv.org/html/2608.05850#bib.bib22)\), orthographic variation, and transliteration\(Saleva,[2020](https://arxiv.org/html/2608.05850#bib.bib29)\)\. More recently, Jochre 3\(Urieli et al\.,[2025](https://arxiv.org/html/2608.05850#bib.bib34)\)substantially improved OCR for printed Yiddish, enabling the digitization of thousands of Yiddish books and supporting the computational use of large historical collections such as the Yiddish Book Center digital library\. In addition, some Yiddish resources now exist within broader multilingual efforts, such as dependency annotation in UD\(Andrews,[2025](https://arxiv.org/html/2608.05850#bib.bib3)\)and domain\-specific NER in EHRI\(Dermentzi & Scheithauer,[2024](https://arxiv.org/html/2608.05850#bib.bib10)\), but no prior work, to our knowledge, has targeted Yiddish through the lens of modern LM development\. Our work aims to fill this gap by combining a dedicated large\-scale pretraining corpus, a broad benchmark for Yiddish language models, and an open\-source Yiddish model\.

## 3Training Data

The quality of pretraining data is a central challenge for low resource LM development\(Kreutzer et al\.,[2022](https://arxiv.org/html/2608.05850#bib.bib21); Doshi et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib11); Ali et al\.,[2025](https://arxiv.org/html/2608.05850#bib.bib2)\)\. This section first examines Yiddish in existing open corpora, focusing on an analysis of the Yiddish portion of mC4 to quantify the extent of noise, machine\-translated content, and misclassified non\-Yiddish text\. It then introducesOytser, our new Yiddish pretraining corpus, which was designed to provide a substantially cleaner and more representative alternative\.

### 3\.1Yiddish in Existing Corpora

We investigate the amount and quality of Yiddish texts in open source corpora\. For that purpose, we focus on mC4\(Xue et al\.,[2021](https://arxiv.org/html/2608.05850#bib.bib38)\), one of the most prominent open multilingual pretraining corpora\. mC4 was introduced as the training corpus for mT5 and is based on filtered Common Crawl data covering 101 languages\. After filtering, the corpus totals 6\.6B pages and 6\.3T tokens\.

Within mC4, approximately 0\.3B words and 0\.1M pages are tagged as Yiddish\. This makes mC4 a potentially attractive default source of Yiddish data for multilingual pretraining\. However, these aggregate statistics provide no information about the linguistic quality of the split itself\.

We therefore carried out an analysis of the Yiddish portion of mC4\.222We ran our analysis on the AI2 implementation of mC4, available at[https://huggingface\.co/datasets/allenai/c4](https://huggingface.co/datasets/allenai/c4)\.To estimate the amount of machine\-translated material, we inspected source URLs in the Yiddish split and identified pages containing Yiddish ISO codes that appeared to be one language\-specific rendering among many automatically generated versions of the same site\. We complemented this with a corpus\-wide pass of an additional language identifier, and with manual assessment of hundreds of domains and sampled pages\. Full technical details of the audit are given in[Appendix D](https://arxiv.org/html/2608.05850#A4)\.

Overall, our analysis suggests that fewer than half \(42\.2%\) of the documents in the Yiddish mC4 split are genuine high\-quality Yiddish from validated native sources\. At least 29\.8% of the documents appear to be machine\-translated, and an additional 21\.9% are Hebrew texts mistakenly identified as Yiddish, and the remaining 6% is a long tail of short fragments, multilingual pages, and small uncatalogued Yiddish sources\.

Table 1:Yiddish in mC4\.These findings show that the existence of a Yiddish split in a large multilingual corpus should not be equated with the availability of native Yiddish pretraining data\. Instead of relying on large automatically constructed Common Crawl corpora, we chose to buildOytseras a genuinely high\-quality and verified alternative, described in the next subsection\.

Table 2:Statistics ofOytser, the Yiddish training corpus by source and type\. The token count is according toMameLoshnLM’s tokenizer\.
### 3\.2Data Sources

We constructedOytser\(אוצר, “treasure”\) a new Yiddish pretraining corpus designed as a higher\-quality alternative to existing open resources\. Our goal was to assemble a corpus that reflects genuine Yiddish usage across domains and registers\. The corpus combines contemporary casual texts with literary materials, reflecting the online footprint of present\-day Yiddish and its much richer textual tradition\. Our corpus is constructed from two source groups, each serving a distinct role:

#### Web\-native Yiddish sources\.

This portion of the data is composed of a wide range of contemporary and diverse forms of Yiddish\. These include Yiddish news websites and magazines, the Yiddish Wikipedia, and a Yiddish translation of the Hebrew Bible; in addition, we included several Yiddish forums, where users write in a freer style that is less well\-conformed with the standard\. In general, the data includes more than 0\.3M native Yiddish documents from various topics, dialects, and communities\.

#### Yiddish Book Center Corpus \(YBC\)

Our largest data source is the Yiddish Digital Library of the YBC,333[https://www\.yiddishbookcenter\.org/collections/digital\-yiddish\-library](https://www.yiddishbookcenter.org/collections/digital-yiddish-library)whose books were OCRed using Jochre3\(Urieli et al\.,[2025](https://arxiv.org/html/2608.05850#bib.bib34)\)\. With more than 12K books and approximately 720M Yiddish words, this is, to our knowledge, the largest and most comprehensive digital source of Yiddish texts currently available\. This source allows us to draw on the long literary history of Yiddish, rather than relying only on its limited contemporary web presence\. Most of the books in this collection date to the last 100 years, making it an important source for capturing the stylistic, dialectal, generic, and historical breadth of Yiddish, spanning not only literary texts but also essays, historical works, and other prose genres\.

Together, these resources provide coverage of multiple genres and registers\.[Table 2](https://arxiv.org/html/2608.05850#S3.T2)summarizes the data sources included in our corpus, together with their genres and sizes\. All in all, the corpus contains more than 915M words from 9 high quality online and literary sources\.

#### Licensing and collection\.

The data used for continued pretraining combines licensed, public\-domain, and publicly accessible Yiddish\-language sources\. The digitized books were used under our research agreement with the Yiddish Book Center, Yiddish Wikipedia under the applicable Wikimedia licenses, Lebns Fragn approved for non\-commercial research, the Forward archive was used for non\-commercial research consistent with the fair\-use provision in its terms of use, and the Yehoash Yiddish translation of the Hebrew Bible from a CC0 public\-domain source\. The remaining web\-native data was collected using source\-specific extraction pipelines from domains already represented in Common Crawl\. Collection was limited to pages accessible without login, subscription, or paywall and followed the applicable site\-level crawling directives at the time of extraction\. Further legal and ethical considerations are discussed in the Ethics Statement\.

#### De\-identification of user\-generated data\.

Before training, we applied a corpus\-wide de\-identification pipeline using carefully designed regular\-expression patterns, consistent with the rule\-based PII filtering approaches used in ROOTS and Dolma\(Laurençon et al\.,[2022](https://arxiv.org/html/2608.05850#bib.bib24); Soldaini et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib32)\)\. These filters were used to remove email addresses, phone numbers, URLs, user handles, and non\-date numerical sequences that could encode identifiers\.

For the Ivelt and Kaveshtiebel forums, we additionally retained only post and comment text and removed all account and interaction metadata, including author usernames, timestamps, signatures, thread identifiers, and reply relations\. The text was then separated from its original discussion context, segmented, and shuffled, so that user profiles, account histories, and conversation graphs were not preserved\. We iteratively sampled and inspected the processed forum data, refining the filters when recurring patterns were identified\. In the final inspected samples, we found no explicit personal information that could readily be linked to a specific user or forum account\.

## 4Kashes: The Yiddish Evaluation Benchmark

For low\-resource languages like Yiddish, the scarcity of evaluation data is as limiting as the scarcity of training data\. Available Yiddish evaluation resources are few and scattered, and to the best of our knowledge, no prior work has systematically benchmarked language models across multiple Yiddish tasks\.

We introduceKashes\(קאַשעס, “difficult questions”\), the first multi\-task evaluation benchmark for Yiddish language models\. The tasks inKashesreflect both the applicative needs of Yiddish scholars and digital humanities researchers, such as linguistic analysis, information extraction, as well as general language model capabilities like translation, commonsense reasoning and paraphrase detection\. In addition to consolidating existing resources, we contribute a new parallel corpus for machine translation – the largest natively authored Yiddish–English benchmark to date\.

TaskDatasetSizeSourceTranslationKashes\-mt5,287\[new task\]see[Section 4\.1](https://arxiv.org/html/2608.05850#S4.SS1)FLORES\+1,012NLLB Team et al\. \([2024](https://arxiv.org/html/2608.05850#bib.bib27)\)POS taggingYiddish Treebank1,079Andrews \([2025](https://arxiv.org/html/2608.05850#bib.bib3)\)Dep\. parsing1,079Transliteration1,079Lemmatization955NEREHRI\-NER4,103Dermentzi & Scheithauer \([2024](https://arxiv.org/html/2608.05850#bib.bib10)\)WikiANN300Rahimi et al\. \([2019](https://arxiv.org/html/2608.05850#bib.bib28)\)newNLP NER1,535Berkovitch & Rusinek \([2021](https://arxiv.org/html/2608.05850#bib.bib5)\)Commonsense QAPIQA625Singh et al\. \([2024](https://arxiv.org/html/2608.05850#bib.bib31)\)Question generationWikiQA293Paraphrase det\.PAWS\-Wiki8000Table 3:Datasets included inKashes\. Size denotes the number of samples \(sentences or sentence pairs\)\.[Table 3](https://arxiv.org/html/2608.05850#S4.T3)summarizes all datasets included inKashes\. The benchmark spans 9 tasks: Four linguistic analysis tasks \(POS tagging, lemmatization, dependency parsing, and transliteration\) are drawn from the YiTB treebank\(Andrews,[2025](https://arxiv.org/html/2608.05850#bib.bib3)\), with structured annotations flattened into sequence\-level formats for generative models\. Named entity recognition is evaluated on three datasets: EHRI\-NER\(Dermentzi & Scheithauer,[2024](https://arxiv.org/html/2608.05850#bib.bib10)\)with annotated Holocaust\-related documents, WikiANN\(Rahimi et al\.,[2019](https://arxiv.org/html/2608.05850#bib.bib28)\)with tagged entities from the 2018 Yiddish Wikipedia dump, and newNLP NER\(Berkovitch & Rusinek,[2021](https://arxiv.org/html/2608.05850#bib.bib5)\)with texts from historical Yiddish newspapers annotated by scholars\.

Three language\-understanding tasks are taken from the Aya machine\-translated collection\(Singh et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib31)\): PIQA, a two\-choice commonsense reasoning task; WikiQA, question generation from a Wikipedia paragraph; and PAWS\-Wiki, paraphrase detection\. Because there are currently no native Yiddish benchmarks for general language understanding, we use these translated tasks to cover evaluation types otherwise unavailable in Yiddish, retaining only those Aya tasks whose translations preserved the essential structure and intent of the original\.

### 4\.1Kashes\-mt: New Yiddish Translation Benchmark

Existing Yiddish–English parallel corpora, such as FLORES\+\(NLLB Team et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib27)\), are derived from English source texts translated into Yiddish\. Therefore, they do not capture the authentic voice of native Yiddish writing\. We address this gap by creating two sentence\-level corpora from bilingual online publications, where the Yiddish texts are originally written by native speakers and the English translations are produced by Yiddish scholars and professional translators\.

We collect human\-translated documents from two web sources: \(1\)\.*Forverts*444[https://forward\.com/yiddish/](https://forward.com/yiddish/)is a digital newspaper containing articles and blog posts from the past decade for which corresponding English versions are available\. Both language versions are authored by native speakers\. \(2\)\.*In geveb*555[https://ingeveb\.org/](https://ingeveb.org/)is a peer\-reviewed, open\-access journal of Yiddish studies that publishes Yiddish literature alongside English translations produced by Yiddish scholars and professional literary translators\.

Yiddish documentsnatively authoredEnglish versionsby scholars & translatorssame articles,two editionsForverts, In gevebDocumentmatching\(TF\-IDF over titles/ shared doc IDs\)Sentencealignment\(SentAlign:LaBSE similarity\)Documentfiltering\(≥\\geq35% aligned\)Sentencefiltering\(score≥0\.65\\geq 0\.65\)De\-duplicationKashes\-mt5,287 pairs6,9526,1765,3585,287aligned sentences:source documents removedfromOytserpretraining corpus

Figure 1:TheKashes\-mt construction pipeline\. Yiddish\-English editions of the same articles are matched and sentence\-aligned to construct a natively authored translation benchmarkWe constructKashes\-mt using a four\-stage pipeline:document matching,sentence alignment,quality filtering, anddeduplication\([Figure 1](https://arxiv.org/html/2608.05850#S4.F1)\)\. ForIn geveb, bilingual editions share document IDs; forForverts, we match similar but nonidentical English titles using TF\-IDF\. SentAlign\(Steingrímsson et al\.,[2023](https://arxiv.org/html/2608.05850#bib.bib33)\)then aligns sentences within matched documents\. We discard documents with fewer than35%35\\%aligned sentences and sentence pairs with a SentAlign score below0\.650\.65,666Score distributions are reported in[Appendix A](https://arxiv.org/html/2608.05850#A1)\.and remove duplicate pairs\. The resulting benchmark contains 995 pairs fromForvertsand 4,292 fromIn geveb, for 5,287 high\-quality pairs in total;[Figure 1](https://arxiv.org/html/2608.05850#S4.F1)reports retention at each stage\. To prevent benchmark leakage, we exclude every source document from theOytserpretraining corpus\(Jacovi et al\.,[2023](https://arxiv.org/html/2608.05850#bib.bib19); Balloccu et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib4)\)\.

## 5MameLoshnLM

We presentMameLoshnLM, the first open\-source large language model for Yiddish, trained on the Yiddish corpus described in[Section 3\.2](https://arxiv.org/html/2608.05850#S3.SS2)\. We name the model*Mame\-Loshn*\(מאַמע־לשון, “mother tongue”\) in reference to the traditional Yiddish term that evokes the language’s intimate connection to home and family\. We show that it outperforms strong baselines of similar scale across a broad set of Yiddish evaluation benchmarks\.

#### Training Details\.

MameLoshnLMwas produced by continued pretraining of Llama\-3\.1\-8B on a Yiddish corpus using a causal language modeling objective\. The model was trained in bfloat16 precision using the 8\-bit AdamW optimizer with a learning rate of2×10−52\\times 10^\{\-5\}, a cosine learning rate scheduler with a 2% warmup ratio, and a weight decay of 0\.01\. Training was conducted for one epoch with a maximum sequence length of 1,024 tokens, a per\-device batch size of 38, and 4 gradient accumulation steps, yielding an effective batch size of approximately 155K tokens\. To mitigate catastrophic forgetting of the base model’s broader capabilities, we include English data from CC100\(Conneau et al\.,[2020](https://arxiv.org/html/2608.05850#bib.bib9)\)during continued pretraining in a Yiddish\-dominant mixture: Yiddish accounts for 72% of words \(92% of tokens\), and English for the remaining 28% \(8% of tokens\)\. We also tested multilingual variants that reallocate part of this non\-Yiddish budget to historically related languages, namely German, Hebrew, Polish and Russian \(See[Section 7\.2](https://arxiv.org/html/2608.05850#S7.SS2)\)\. In total, the model was trained on approximately 5\.7 billion tokens over 36,663 optimization steps\. Gradient checkpointing was enabled to reduce memory consumption\. Training was conducted on a single NVIDIA H200 GPU for approximately 207 GPU\-hours\.

## 6Experimental Setup

We evaluateMameLoshnLMalongside five open\-weight baseline models of similar scale: Llama 3\.1 8B\(Grattafiori et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib17)\), Qwen3 8B\(Yang et al\.,[2025](https://arxiv.org/html/2608.05850#bib.bib39)\), BLOOMZ 7B\(Muennighoff et al\.,[2022](https://arxiv.org/html/2608.05850#bib.bib26)\), Gemma\-2 9B\(Gemma Team,[2024](https://arxiv.org/html/2608.05850#bib.bib15)\), and EuroLLM 9B\(Martins et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib25)\)\. Qwen3 is the only baseline whose documentation explicitly lists Yiddish as a supported language\. The remaining models do not claim explicit Yiddish support, but provide strong, comparable\-size open\-weight baselines with varying degrees of multilingual capability; Llama 3\.1 additionally provides the direct comparison toMameLoshnLM’s base model\. To our knowledge, none has previously been systematically evaluated across a broad Yiddish benchmark\. We evaluate all models under identical 1\-, 3\-, and 5\-shot conditions, sampling demonstrations with the same random seed in a leave\-one\-out scheme\. Example inputs and demonstrations are provided in[Appendix C](https://arxiv.org/html/2608.05850#A3)\.

## 7Results

### 7\.1Performance onKashes

We report 5\-shot results across all tasks and models in[Table 4](https://arxiv.org/html/2608.05850#S7.T4), with full results for all shot settings \(1, 3, and 5\-shot\) in[Appendix B](https://arxiv.org/html/2608.05850#A2)\.

Across 14 evaluations,MameLoshnLMachieves the best overall performance, with the highest average score \(62\.6\), ahead of Gemma\-2 9B \(57\.0\), Llama 3\.1 8B \(56\.8\), and Qwen3 8B \(54\.7\)\. Its gains are concentrated on Yiddish\-centered tasks: it leads on POS tagging, dependency parsing, transliteration, EHRI and newNLP NER\.

Table 4:Evaluation results onKashesbenchmark\.Bold= best model per row\. All metrics: higher is better\.It also remains competitive on the remaining tasks, including near\-best performance on WikiANN NER and WikiQA\. The main exceptions are PAWS\-Wiki and PIQA, which are Aya benchmarks based on machine\-translated versions of widely used multilingual datasets\. The performance ofMameLoshnLMis especially strong in the English to Yiddish translation \(more than 11 COMET points improvement over the closest competitor\)\. It indicates that continued training enabled the generation of fluent texts in Yiddish\. We further support this claim by qualitative analysis of sentences translated into Yiddish in[Section 8](https://arxiv.org/html/2608.05850#S8)\.

Overall, the results indicate that continuing pretraining in authentic Yiddish data yields substantial improvements in tasks that require lexical, orthographic, and syntactic command of Yiddish, while maintaining competitive performance on more general benchmarks\.

### 7\.2Effect of related\-language mixing\.

A practical question in low\-resource continued pretraining is how to use the non\-target\-language portion of the training budget\(cf\. Koto et al\.,[2025](https://arxiv.org/html/2608.05850#bib.bib20)\)\. For Yiddish, a natural hypothesis is that historically related languages such as Hebrew and German may be more useful than English, either because they share lexical material with Yiddish or because they may support transfer on linguistically relevant structures\. We test this directly by keeping the same Yiddish corpus and continued\-pretraining setup asMameLoshnLM, and training two variants that modify only the remaining mixture\. Data for each language was drawn from CC100\(Conneau et al\.,[2020](https://arxiv.org/html/2608.05850#bib.bib9)\)in fixed\-size blocks\. The two configurations were designed to test whether reallocating part of the English budget to related languages \(Rebalanced\) or supplementing it with additional multilingual data \(Expanded\) would improve Yiddish adaptation\. InRebalanced, part of the English budget is redistributed to Hebrew and German; inExpanded, we add Hebrew, German, Russian, and Polish while keeping the absolute English amount \(exact mixture proportions in[Appendix F](https://arxiv.org/html/2608.05850#A6)\)\. Because Yiddish tokenizes much more densely than the other languages, these word\-level changes reduce the Yiddish share of training tokens from 91\.8% inMameLoshnLMto 85\.2% and 71\.0%, respectively\. Although the multilingual variants remain competitive in translation and show isolated gains, they do not yield a consistently better Yiddish model, and on several Yiddish\-centered tasks they underperform not onlyMameLoshnLMbut even the base Llama 3\.1 model \(See[Appendix F](https://arxiv.org/html/2608.05850#A6)for full results\)\. These findings suggest that, in our setting, maintaining a strongly Yiddish\-dominant training signal with a limited amount of English is a more effective strategy than reallocating that budget to related languages\.

## 8Analysis

Yiddish is a useful test case for analyzing what multilingual language models learn in low\-resource settings\. A general multilingual model can often produce something understandable in Yiddish because Yiddish is typically represented, at least to some extent, in the training data, and also overlaps with neighboring languages and shares script with Hebrew\. But this surface fluency can be misleading: models may produce plausible Yiddish while still missing many of the features that make the language sound native\. This makes Yiddish a useful case for a broader question: when multilingual models appear to handle a low\-resource language, are they learning the language itself, or a flatter approximation shaped by noisy data?

ProbeMameLoshnLMLlama 3\.1 8BGoldppLK vocabulary \(Eng→\\rightarrowYid translation\)LK content word rate \(%\)4\.71\.66\.2<10−161<10^\{\-161\}LK sentence match rate \(%\)52\.416\.0—<10−229<10^\{\-229\}Morphology \(lemmatization change accuracy\)ge\-participles \(nn=236\)50\.85\.1—<10−29<10^\{\-29\}Hebrew\-origin plurals \(nn=53\)30\.22\.3—<0\.001<0\.001Determiner paradigm \(nn=291\)28\.215\.5—<10−5<10^\{\-5\}Auxiliary control pairזײַן “to be” — suppletive \(nn=460\)20\.08\.9—<10−9<10^\{\-9\}האָבן “to have” — regular \(nn=163\)84\.784\.7—1\.0 \(n\.s\.\)Table 5:Linguistic competence probes,MameLoshnLMvs\. Llama 3\.1 8B \(5\-shot\)\. Translation metrics are computed overKashes\-mt sentence pairs; morphological metrics are lemmatization change accuracy on the UD Yiddish\-YiTB test set\. Gold denotes the rate in native references\. Full results in[Table 12](https://arxiv.org/html/2608.05850#A5.T12)\.We examine this question directly by probing phenomena that distinguish native Yiddish from such an approximation\. Our analysis777For reproducibility, implementation details for all analysis steps are provided in[Appendix E](https://arxiv.org/html/2608.05850#A5)\.focuses on two cases: the lexical layer that is derived from Hebrew/Aramaic \(known as*loshn\-koydesh*, abbreviated hereafter LK\) and Yiddish\-specific morphology\. Across both, general multilingual models often produce an intelligible but systematically reduced variety of Yiddish, whileMameLoshnLMmore closely matches patterns found in native Yiddish data\.

#### General multilingual models underproduce core*loshn\-koydesh*vocabulary\.

A defining property of Yiddish is its substantial Hebrew lexical layer, which includes many frequent everyday words\. If a model has learned native\-like Yiddish, it should produce these items naturally\. If instead it has only shallow or noisy knowledge of Yiddish, these words should be systematically underproduced, avoided, or replaced\. We test this directly in English→\\rightarrowYiddish translation of theKashes\-mt benchmark using the loshn\-koydesh pronunciation lexicon888Based on Niborski’s Lexicon of Loshn\-Koydesh words,[https://github\.com/ibleaman/loshn\-koydesh\-pronunciation](https://github.com/ibleaman/loshn-koydesh-pronunciation)\.to identify LK words\. In 5,287 5\-shot translations, gold references contain LK words in 6\.2% of content tokens\.MameLoshnLMproduces LK words at 4\.7%, whereas Llama 3\.1 8B produces only 1\.6% \(p<10−229p<10^\{\-229\}, paired t\-test\)\. The same pattern appears at the sentence level: when a gold translation contains an LK word,MameLoshnLMis much more likely than Llama to produce a matching LK item in its translation of the same input \(52\.4% vs\. 16\.0% in 5\-shot\)\. Importantly, this effect is not simply a reflection of overall translation quality\. Per\-sentence LK recall correlates only weakly with COMET score \(Spearmanρ=0\.23\\rho=0\.23,p<0\.001p<0\.001\), suggesting that COMET is largely blind to loss of the*loshn\-koydesh*lexical layer\. A translation can therefore score well on overall quality while replacing Hebrew\-origin vocabulary with Germanic alternatives\. This shows that Llama misses many LK words that appear in native Yiddish renderings of the same content\. Moreover, Llama’s LK output is concentrated in proper nouns and culturally salient items such as ישׂראל \(Israel\), רבי \(Rebbe\), and תּורה \(Torah\), rather than in the common vocabulary that characterizes natural Yiddish usage\. See Appendix[Table 13](https://arxiv.org/html/2608.05850#A5.T13)for the most frequent LK words and their frequencies in both models\.

This deficit is not random\. When the reference contains an LK noun with a Germanic alternative, Llama often substitutes the Germanic form, e\.g\., משפּחה “family” with פֿאַמיליע, מלחמה “war” with קריג, and פּנים “face” with געזיכט\. For frequent LK function words with no simple Germanic equivalent, such as אפֿשר “perhaps,” כּמעט “almost,” and כּדי “in order to,” Llama often avoids the lexical item through paraphrase or restructuring\. The result is often understandable, but it is noticeably more Germanized and less native\-like than either the reference or the output ofMameLoshnLM\. The contrast therefore concerns not only how much Yiddish the models produce, but what kind of Yiddish they produce\.

#### The same non\-native pattern appears in Yiddish\-specific morphology\.

If general multilingual models have only limited command of Yiddish\-specific morphology, they should also fail on irregular inflectional patterns, where the correct lemma cannot be recovered by simple surface copying\. This is exactly what we observe in lemmatization\. The strongest contrast appears on past participles, normally formed with age\-prefix but rife with irregularities\. On 236 such tokens,MameLoshnLMreaches 50\.8% accuracy, while Llama reaches only 5\.1% \(p<10−29p<10^\{\-29\}\)\. Similar gaps appear on Hebrew\-origin plural patterns such as “tales” מעשׂיות→\\rightarrowמעשׂה \(30\.2% vs\. 2\.3%\), and on the definite article paradigm, where surface forms such as די, דאָס, and דעם, all corresponding to Englishthe, must be mapped to the same citation lemma דער \(28\.2% vs\. 15\.5%\)\.

The same contrast appears within auxiliary verbs\. On regular forms of האָבן, such as האָט→\\rightarrowהאָבן, the two models perform identically \(84\.7% vs\. 84\.7%\)\. This suggests that both can handle cases where the lemma remains locally recoverable from the surface form\. But on forms of the highly irregular verb זײַן “to be,” whose inflected forms often look very different from the lemma,MameLoshnLMperforms much better than Llama \(20\.0% vs\. 8\.9%\)\. In these cases, Llama’s dominant strategy is to copy the input unchanged, suggesting that it often fails to recover the underlying paradigm\. The gap is therefore not uniform across morphology: it is largest when successful lemmatization requires knowledge of lexeme\-specific inflectional system rather than simple regular transformations\.

#### This pattern is consistent with the data\-quality picture in[Section 3\.1](https://arxiv.org/html/2608.05850#S3.SS1)\.

As discussed above, publicly available Yiddish web corpora contain substantial amounts of machine\-translated material, weakening the already limited signal available for learning native\-like Yiddish\. Our mC4 audit provides a direct measurement of this effect: machine\-translated pages in the Yiddish split show an LK rate of 3\.6%, less than half the 10\.2% of validated native sources \([Appendix D](https://arxiv.org/html/2608.05850#A4)\)\. Llama’s own LK production rate falls below even that of the machine\-translated web text\. In generation, the model reproduces the same lexical depletion that characterizes its likely training data\. General multilingual models can thus produce superficially fluent Yiddish while missing the lexical and morphological features that make the language sound native\. Continued pretraining on curated Yiddish text reduces this gap and recovers much of the missing competence\. More broadly, this suggests a problem that may extend to other low\-resource languages with limited web presence: multilingual web corpora can support surface\-level fluency while still weakening language\-specific properties that matter for authentic generation\.

## 9Discussion and Conclusion

A natural question is why a language like Yiddish needs a dedicated language model\. Yiddish is spoken by roughly one million people worldwide, and its textual heritage is vast and increasingly the subject of active research\. Yet without adequate language technologies, this heritage remains difficult to search, organize, and analyze at scale\. Notably, several datasets inKashes, such as EHRI\-NER and newNLP NER, were created by digital humanities scholars whose primary goal is extracting information from Holocaust testimonies and historical newspapers, not evaluating language models\. Improved performance on these tasks therefore reflects not just benchmark gains but progress toward tools that these communities can actually use\.

Beyond its practical importance, Yiddish is an informative case for multilingual NLP\. Its position at the intersection of several language families, combined with its unique script, makes it a compelling testbed for research on tokenization, cross\-lingual transfer, multilingual data mixing, and the effects of data scarcity and quality on language model development\. This problem likely extends to other languages whose authentic usage diverges from their web footprint, including languages in diglossia situations and those with fragmented online presence\.

This work presents the first comprehensive LLM development effort for Yiddish, encompassing corpus construction, benchmark curation, model training, and evaluation\. Building on this foundation, future work can pursue instruction tuning for interactive use, the creation of additional training and evaluation resources, and the application ofMameLoshnLMto large\-scale digital humanities workflows\. We hope that the work presented here, includingMameLoshnLM,Oytser, andKashes, will contribute to the Yiddish\-speaking and research communities and provide a useful reference point for similar efforts in other underrepresented languages\.

## Acknowledgments

The authors would like to thank the Yiddish Book Center for its permission to use the Steven Spielberg Digital Yiddish Library, and*In geveb: A Journal of Yiddish Studies*for granting permission to use its expert translation materials as part of our evaluation benchmark\. OG research is funded by the Blavatnik Family Foundation, his work on this paper was also partially funded by UniDive COST Action \(\#CA21167\)\.

## Ethics Statement

#### Purpose and institutional partnerships\.

Yiddish is a language of significant historical and cultural importance that remains severely underrepresented in modern NLP resources\. This work aims to support Yiddish\-language preservation, accessibility, and research, and was developed in partnership with institutions dedicated to Yiddish culture\. The Yiddish Book Center licensed the Steven Spielberg Digital Yiddish Library for model training, andIn geveblicensed material used in our evaluation benchmark\. MameLoshnLM is released under a non\-commercial license consistent with the terms of these agreements\.

#### Copyright and fair\-use considerations\.

The corpus combines material used under institutional agreements, open and public\-domain licenses, applicable source terms, and publicly accessible web text\. More than 82% of the training tokens derive from sources covered by the first three categories\. The web\-native material was collected for non\-commercial academic research exclusively from pages accessible without login, subscription, paywall, or circumvention of technical access controls, and in accordance with applicable site\-level crawling directives\. We verified that these domains were already represented in Common Crawl and in widely used multilingual corpora derived from it, including mC4 and OSCAR\(Xue et al\.,[2021](https://arxiv.org/html/2608.05850#bib.bib38); Abadji et al\.,[2022](https://arxiv.org/html/2608.05850#bib.bib1)\)\. Our source\-specific pipelines therefore provide a cleaner and more complete extraction of Yiddish text from established public\-web sources rather than exposing previously inaccessible material\.

Consistent with the considerations applied in Dolma\(Soldaini et al\.,[2024](https://arxiv.org/html/2608.05850#bib.bib32)\)to training on publicly available web data, we rely on fair use and analogous research exceptions as the legal basis for this training\-stage use, which is non\-expressive and transformative: a computational use intended to learn general linguistic patterns rather than to provide access to or substitute for individual works\. This position is supported by the treatment of non\-commercial research and analysis under United States fair\-use principles and by the Israeli Ministry of Justice’s conclusion that machine\-learning uses will generally fall within fair\-use and related statutory exceptions\(U\.S\. Copyright Office,[2025](https://arxiv.org/html/2608.05850#bib.bib35); Israel Ministry of Justice,[2022](https://arxiv.org/html/2608.05850#bib.bib18)\)\.

#### Personal information and privacy\.

To our knowledge, Ivelt and Kaveshtiebel are the only substantial publicly accessible digital sources we identified for contemporary, informal, community\-authored Yiddish\. Their inclusion was important for representing everyday language that is largely absent from historical and edited collections\. They are also already prominent in Common Crawl\-derived training data: in the Yiddish portion of mC4, Ivelt is the most frequent source domain and Kaveshtiebel is the fifth most frequent\. Following the harm\-minimization approach recommended by the Association of Internet Researchers\(Franzke et al\.,[2020](https://arxiv.org/html/2608.05850#bib.bib14)\), we applied source\-specific safeguards beyond the corpus\-wide processing described in Section 3\.2\. We removed account and interaction structure, retained only decontextualized post and comment text, and iteratively inspected processed samples for identifying information, refining the procedure whenever recurring risks were found\. No explicit personal information readily linkable to a specific user or account was found in the final inspected samples\. These measures were intended to preserve the linguistic value of the data while minimizing disclosure and re\-identification risks\.

## References

- Abadji et al\. \(2022\)Julien Abadji, Pedro Ortiz Suarez, Laurent Romary, and Benoît Sagot\.Towards a cleaner document\-oriented multilingual crawled corpus\.In Nicoletta Calzolari, Frédéric Béchet, Philippe Blache, Khalid Choukri, Christopher Cieri, Thierry Declerck, Sara Goggi, Hitoshi Isahara, Bente Maegaard, Joseph Mariani, Hélène Mazo, Jan Odijk, and Stelios Piperidis \(eds\.\),*Proceedings of the Thirteenth Language Resources and Evaluation Conference*, pp\. 4344–4355, Marseille, France, June 2022\. European Language Resources Association\.URL[https://aclanthology\.org/2022\.lrec\-1\.463/](https://aclanthology.org/2022.lrec-1.463/)\.
- Ali et al\. \(2025\)Mehdi Ali, Manuel Brack, Max Lübbering, Elias Wendt, Abbas Goher Khan, Richard Rutmann, Alex Jude, Maurice Kraus, Alexander Arno Weber, Felix Stollenwerk, David Kaczér, Florian Mai, Lucie Flek, Rafet Sifa, Nicolas Flores\-Herr, Joachim Koehler, Patrick Schramowski, Michael Fromm, and Kristian Kersting\.Judging quality across languages: A multilingual approach to pretraining data filtering with language models\.In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng \(eds\.\),*Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing*, pp\. 8870–8909, Suzhou, China, November 2025\. Association for Computational Linguistics\.ISBN 979\-8\-89176\-332\-6\.doi:10\.18653/v1/2025\.emnlp\-main\.449\.URL[https://aclanthology\.org/2025\.emnlp\-main\.449/](https://aclanthology.org/2025.emnlp-main.449/)\.
- Andrews \(2025\)Matthew Kirk Andrews\.YiTB: the yiddish tree bank, 2025\.URL[https://github\.com/UniversalDependencies/UD\_Yiddish\-YiTB](https://github.com/UniversalDependencies/UD_Yiddish-YiTB)\.
- Balloccu et al\. \(2024\)Simone Balloccu, Patrícia Schmidtová, Mateusz Lango, and Ondrej Dusek\.Leak, cheat, repeat: Data contamination and evaluation malpractices in closed\-source LLMs\.In Yvette Graham and Matthew Purver \(eds\.\),*Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pp\. 67–93, St\. Julian’s, Malta, March 2024\. Association for Computational Linguistics\.doi:10\.18653/v1/2024\.eacl\-long\.5\.URL[https://aclanthology\.org/2024\.eacl\-long\.5/](https://aclanthology.org/2024.eacl-long.5/)\.
- Berkovitch & Rusinek \(2021\)Ephraim Berkovitch and Sinai Rusinek\.New Languages for NLP: Yiddish — annotated data and models\.GitHub repository, New Languages for NLP project, Princeton University, 2021\.URL[https://github\.com/New\-Languages\-for\-NLP/yiddish](https://github.com/New-Languages-for-NLP/yiddish)\.MIT License\.
- Bleaman & Nove \(2025\)Isaac L Bleaman and Chaya R Nove\.The corpus of spoken yiddish in europe: Goals, methods, and applications\.*Language Documentation & Conservation*, 19, 2025\.
- Brown & Marivate \(2025\)Nathan Brown and Vukosi Marivate\.Pula: Training large language models for setswana\.In Luis Chiruzzo, Alan Ritter, and Lu Wang \(eds\.\),*Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\)*, pp\. 6634–6656, Albuquerque, New Mexico, April 2025\. Association for Computational Linguistics\.ISBN 979\-8\-89176\-189\-6\.doi:10\.18653/v1/2025\.naacl\-long\.338\.URL[https://aclanthology\.org/2025\.naacl\-long\.338/](https://aclanthology.org/2025.naacl-long.338/)\.
- Ćavar et al\. \(2016\)Malgorzata Ćavar, Damir Ćavar, Dov\-Ber Kerler, and Anya Quilitzsch\.Generating a Yiddish speech corpus, forced aligner and basic ASR system for the AHEYM project\.In Nicoletta Calzolari, Khalid Choukri, Thierry Declerck, Sara Goggi, Marko Grobelnik, Bente Maegaard, Joseph Mariani, Helene Mazo, Asuncion Moreno, Jan Odijk, and Stelios Piperidis \(eds\.\),*Proceedings of the Tenth International Conference on Language Resources and Evaluation \(LREC’16\)*, pp\. 4688–4693, Portorož, Slovenia, May 2016\. European Language Resources Association \(ELRA\)\.URL[https://aclanthology\.org/L16\-1744/](https://aclanthology.org/L16-1744/)\.
- Conneau et al\. \(2020\)Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov\.Unsupervised cross\-lingual representation learning at scale\.In Dan Jurafsky, Joyce Chai, Natalie Schluter, and Joel Tetreault \(eds\.\),*Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics*, pp\. 8440–8451, Online, July 2020\. Association for Computational Linguistics\.doi:10\.18653/v1/2020\.acl\-main\.747\.URL[https://aclanthology\.org/2020\.acl\-main\.747/](https://aclanthology.org/2020.acl-main.747/)\.
- Dermentzi & Scheithauer \(2024\)Maria Dermentzi and Hugo Scheithauer\.Repurposing holocaust\-related digital scholarly editions to develop multilingual domain\-specific named entity recognition tools\.In Isuri Anuradha, Martin Wynne, Francesca Frontini, and Alistair Plum \(eds\.\),*Proceedings of the First Workshop on Holocaust Testimonies as Language Resources \(HTRes\) @ LREC\-COLING 2024*, pp\. 18–28, Torino, Italia, May 2024\. ELRA and ICCL\.URL[https://aclanthology\.org/2024\.htres\-1\.3/](https://aclanthology.org/2024.htres-1.3/)\.
- Doshi et al\. \(2024\)Meet Doshi, Raj Dabre, and Pushpak Bhattacharyya\.Pretraining language models using translationese\.In Yaser Al\-Onaizan, Mohit Bansal, and Yun\-Nung Chen \(eds\.\),*Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing*, pp\. 5843–5862, Miami, Florida, USA, November 2024\. Association for Computational Linguistics\.doi:10\.18653/v1/2024\.emnlp\-main\.334\.URL[https://aclanthology\.org/2024\.emnlp\-main\.334/](https://aclanthology.org/2024.emnlp-main.334/)\.
- Elhady et al\. \(2025\)Ahmed Elhady, Eneko Agirre, and Mikel Artetxe\.Emergent abilities of large language models under continued pre\-training for language adaptation\.In Wanxiang Che, Joyce Nabende, Ekaterina Shutova, and Mohammad Taher Pilehvar \(eds\.\),*Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pp\. 32174–32186, Vienna, Austria, July 2025\. Association for Computational Linguistics\.ISBN 979\-8\-89176\-251\-0\.doi:10\.18653/v1/2025\.acl\-long\.1547\.URL[https://aclanthology\.org/2025\.acl\-long\.1547/](https://aclanthology.org/2025.acl-long.1547/)\.
- Etxaniz et al\. \(2024\)Julen Etxaniz, Oscar Sainz, Naiara Perez, Itziar Aldabe, German Rigau, Eneko Agirre, Aitor Ormazabal, Mikel Artetxe, and Aitor Soroa\.Latxa: An open language model and evaluation suite for Basque\.In Lun\-Wei Ku, Andre Martins, and Vivek Srikumar \(eds\.\),*Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pp\. 14952–14972, Bangkok, Thailand, August 2024\. Association for Computational Linguistics\.doi:10\.18653/v1/2024\.acl\-long\.799\.URL[https://aclanthology\.org/2024\.acl\-long\.799/](https://aclanthology.org/2024.acl-long.799/)\.
- Franzke et al\. \(2020\)Aline Shakti Franzke, Anja Bechmann, Charles Melvin Ess, and Michael Zimmer \(eds\.\)\.*Internet Research: Ethical Guidelines 3\.0*, volume 3\.AoIR \(The International Association of Internet Researchers\), 2020\.
- Gemma Team \(2024\)Gemma Team\.Gemma 2: Improving open language models at a practical size, 2024\.URL[https://arxiv\.org/abs/2408\.00118](https://arxiv.org/abs/2408.00118)\.
- Genzel et al\. \(2009\)Dmitriy Genzel, Klaus Macherey, and Jakob Uszkoreit\.Creating a high\-quality machine translation system for a low\-resource language: Yiddish\.In*Proceedings of Machine Translation Summit XII: Papers*, Ottawa, Canada, August 26\-30 2009\.URL[https://aclanthology\.org/2009\.mtsummit\-papers\.4/](https://aclanthology.org/2009.mtsummit-papers.4/)\.
- Grattafiori et al\. \(2024\)Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al\-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al\.The llama 3 herd of models\.*arXiv preprint arXiv:2407\.21783*, 2024\.
- Israel Ministry of Justice \(2022\)Israel Ministry of Justice\.Uses of copyrighted materials for machine learning\.Technical report, Ministry of Justice, Jerusalem, Israel, 12 2022\.URL[https://www\.gov\.il/BlobFolder/legalinfo/machine\-learning/he/18\-12\-2022\.pdf](https://www.gov.il/BlobFolder/legalinfo/machine-learning/he/18-12-2022.pdf)\.Non\-binding legal opinion on the permissibility of machine learning training under the Israeli Copyright Act, 2007\.
- Jacovi et al\. \(2023\)Alon Jacovi, Avi Caciularu, Omer Goldman, and Yoav Goldberg\.Stop uploading test data in plain text: Practical strategies for mitigating data contamination by evaluation benchmarks\.In Houda Bouamor, Juan Pino, and Kalika Bali \(eds\.\),*Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing*, pp\. 5075–5084, Singapore, December 2023\. Association for Computational Linguistics\.doi:10\.18653/v1/2023\.emnlp\-main\.308\.URL[https://aclanthology\.org/2023\.emnlp\-main\.308/](https://aclanthology.org/2023.emnlp-main.308/)\.
- Koto et al\. \(2025\)Fajri Koto, Rituraj Joshi, Nurdaulet Mukhituly, Yuxia Wang, Zhuohan Xie, Rahul Pal, Daniil Orel, Parvez Mullah, Diana Turmakhan, Maiya Goloburda, Mohammed Kamran, Samujjwal Ghosh, Bokang Jia, Jonibek Mansurov, Mukhammed Togmanov, Debopriyo Banerjee, Nurkhan Laiyk, Akhmed Sakip, Xudong Han, Ekaterina Kochmar, Alham Fikri Aji, Aaryamonvikram Singh, Alok Anil Jadhav, Satheesh Katipomu, Samta Kamboj, Monojit Choudhury, Gurpreet Gosal, Gokulakrishnan Ramakrishnan, Biswajit Mishra, Sarath Chandran, Avraham Sheinin, Natalia Vassilieva, Neha Sengupta, and Preslav Nakov\.Sherkala\-chat: Building a state\-of\-the\-art LLM for kazakh in a moderately resourced setting\.In*Second Conference on Language Modeling*, 2025\.URL[https://openreview\.net/forum?id=wRcTCcb0H5](https://openreview.net/forum?id=wRcTCcb0H5)\.
- Kreutzer et al\. \(2022\)Julia Kreutzer, Isaac Caswell, Lisa Wang, Ahsan Wahab, Daan van Esch, Nasanbayar Ulzii\-Orshikh, Allahsera Tapo, Nishant Subramani, Artem Sokolov, Claytone Sikasote, Monang Setyawan, Supheakmungkol Sarin, Sokhar Samb, Benoît Sagot, Clara Rivera, Annette Rios, Isabel Papadimitriou, Salomey Osei, Pedro Ortiz Suarez, Iroro Orife, Kelechi Ogueji, Andre Niyongabo Rubungo, Toan Q\. Nguyen, Mathias Müller, André Müller, Shamsuddeen Hassan Muhammad, Nanda Muhammad, Ayanda Mnyakeni, Jamshidbek Mirzakhalov, Tapiwanashe Matangira, Colin Leong, Nze Lawson, Sneha Kudugunta, Yacine Jernite, Mathias Jenny, Orhan Firat, Bonaventure F\. P\. Dossou, Sakhile Dlamini, Nisansa de Silva, Sakine Çabuk Ballı, Stella Biderman, Alessia Battisti, Ahmed Baruwa, Ankur Bapna, Pallavi Baljekar, Israel Abebe Azime, Ayodele Awokoya, Duygu Ataman, Orevaoghene Ahia, Oghenefego Ahia, Sweta Agrawal, and Mofetoluwa Adeyemi\.Quality at a glance: An audit of web\-crawled multilingual datasets\.*Transactions of the Association for Computational Linguistics*, 10:50–72, 2022\.doi:10\.1162/tacl\_a\_00447\.URL[https://aclanthology\.org/2022\.tacl\-1\.4/](https://aclanthology.org/2022.tacl-1.4/)\.
- Kulick et al\. \(2022\)Seth Kulick, Neville Ryant, Beatrice Santorini, Joel Wallenberg, and Assaf Urieli\.A part\-of\-speech tagger for yiddish\.*arXiv preprint arXiv:2204\.01175*, 2022\.
- Kuulmets et al\. \(2024\)Hele\-Andra Kuulmets, Taido Purason, Agnes Luhtaru, and Mark Fishel\.Teaching llama a new language through cross\-lingual knowledge transfer\.In Kevin Duh, Helena Gomez, and Steven Bethard \(eds\.\),*Findings of the Association for Computational Linguistics: NAACL 2024*, pp\. 3309–3325, Mexico City, Mexico, June 2024\. Association for Computational Linguistics\.doi:10\.18653/v1/2024\.findings\-naacl\.210\.URL[https://aclanthology\.org/2024\.findings\-naacl\.210/](https://aclanthology.org/2024.findings-naacl.210/)\.
- Laurençon et al\. \(2022\)Hugo Laurençon, Lucile Saulnier, Thomas Wang, Christopher Akiki, Albert Villanova del Moral, Teven Le Scao, Leandro Von Werra, Chenghao Mou, Eduardo González Ponferrada, Huu Nguyen, Jörg Frohberg, Mario Šaško, Quentin Lhoest, Angelina McMillan\-Major, Gérard Dupont, Stella Biderman, Anna Rogers, Loubna Ben allal, Francesco De Toni, Giada Pistilli, Olivier Nguyen, Somaieh Nikpoor, Maraim Masoud, Pierre Colombo, Javier de la Rosa, Paulo Villegas, Tristan Thrush, Shayne Longpre, Sebastian Nagel, Leon Weber, Manuel Romero Muñoz, Jian Zhu, Daniel Van Strien, Zaid Alyafeai, Khalid Almubarak, Vu Minh Chien, Itziar Gonzalez\-Dios, Aitor Soroa, Kyle Lo, Manan Dey, Pedro Ortiz Suarez, Aaron Gokaslan, Shamik Bose, David Ifeoluwa Adelani, Long Phan, Hieu Tran, Ian Yu, Suhas Pai, Jenny Chim, Violette Lepercq, Suzana Ilic, Margaret Mitchell, Sasha Luccioni, and Yacine Jernite\.The bigscience ROOTS corpus: A 1\.6TB composite multilingual dataset\.In*Thirty\-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track*, 2022\.URL[https://openreview\.net/forum?id=UoEw6KigkUn](https://openreview.net/forum?id=UoEw6KigkUn)\.
- Martins et al\. \(2024\)Pedro Henrique Martins, Patrick Fernandes, João Alves, Nuno M\. Guerreiro, Ricardo Rei, Duarte M\. Alves, José Pombal, Amin Farajian, Manuel Faysse, Mateusz Klimaszewski, Pierre Colombo, Barry Haddow, José G\. C\. de Souza, Alexandra Birch, and André F\. T\. Martins\.Eurollm: Multilingual language models for europe, 2024\.URL[https://arxiv\.org/abs/2409\.16235](https://arxiv.org/abs/2409.16235)\.
- Muennighoff et al\. \(2022\)Niklas Muennighoff, Thomas Wang, Lintang Sutawika, Adam Roberts, Stella Biderman, Teven Le Scao, M Saiful Bari, Sheng Shen, Zheng\-Xin Yong, Hailey Schoelkopf, et al\.Crosslingual generalization through multitask finetuning\.*arXiv preprint arXiv:2211\.01786*, 2022\.
- NLLB Team et al\. \(2024\)NLLB Team, Marta R\. Costa\-jussà, James Cross, Onur Çelebi, Maha Elbayad, Kenneth Heafield, Kevin Heffernan, Elahe Kalbassi, Janice Lam, Daniel Licht, Jean Maillard, Anna Sun, Skyler Wang, Guillaume Wenzek, Al Youngblood, Bapi Akula, Loic Barrault, Gabriel Mejia Gonzalez, Prangthip Hansanti, John Hoffman, Semarley Jarrett, Kaushik Ram Sadagopan, Dirk Rowe, Shannon Spruit, Chau Tran, Pierre Andrews, Necip Fazil Ayan, Shruti Bhosale, Sergey Edunov, Angela Fan, Cynthia Gao, Vedanuj Goswami, Francisco Guzmán, Philipp Koehn, Alexandre Mourachko, Christophe Ropers, Safiyyah Saleem, Holger Schwenk, and Jeff Wang\.Scaling neural machine translation to 200 languages\.*Nature*, 630\(8018\):841–846, 2024\.ISSN 1476\-4687\.doi:10\.1038/s41586\-024\-07335\-x\.URL[https://doi\.org/10\.1038/s41586\-024\-07335\-x](https://doi.org/10.1038/s41586-024-07335-x)\.
- Rahimi et al\. \(2019\)Afshin Rahimi, Yuan Li, and Trevor Cohn\.Massively multilingual transfer for NER\.In Anna Korhonen, David Traum, and Lluís Màrquez \(eds\.\),*Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics*, pp\. 151–164, Florence, Italy, July 2019\. Association for Computational Linguistics\.doi:10\.18653/v1/P19\-1015\.URL[https://aclanthology\.org/P19\-1015/](https://aclanthology.org/P19-1015/)\.
- Saleva \(2020\)Jonne Saleva\.A multi\-orthography parallel corpus of Yiddish nouns\.In Nicoletta Calzolari, Frédéric Béchet, Philippe Blache, Khalid Choukri, Christopher Cieri, Thierry Declerck, Sara Goggi, Hitoshi Isahara, Bente Maegaard, Joseph Mariani, Hélène Mazo, Asuncion Moreno, Jan Odijk, and Stelios Piperidis \(eds\.\),*Proceedings of the Twelfth Language Resources and Evaluation Conference*, pp\. 948–952, Marseille, France, May 2020\. European Language Resources Association\.ISBN 979\-10\-95546\-34\-4\.URL[https://aclanthology\.org/2020\.lrec\-1\.119/](https://aclanthology.org/2020.lrec-1.119/)\.
- Santorini \(2021\)Beatrice Santorini\.Penn parsed corpus of historical yiddish, v1\.0, 2021\.URL[https://github\.com/beatrice57/penn\-parsed\-corpus\-of\-historical\-yiddish](https://github.com/beatrice57/penn-parsed-corpus-of-historical-yiddish)\.Accessed: 2026\-03\-26\.
- Singh et al\. \(2024\)Shivalika Singh, Freddie Vargus, Daniel D’souza, Börje F\. Karlsson, Abinaya Mahendiran, Wei\-Yin Ko, Herumb Shandilya, Jay Patel, Deividas Mataciunas, Laura O’Mahony, Mike Zhang, Ramith Hettiarachchi, Joseph Wilson, Marina Machado, Luisa Moura, Dominik Krzemiński, Hakimeh Fadaei, Irem Ergun, Ifeoma Okoh, Aisha Alaagib, Oshan Mudannayake, Zaid Alyafeai, Vu Chien, Sebastian Ruder, Surya Guthikonda, Emad Alghamdi, Sebastian Gehrmann, Niklas Muennighoff, Max Bartolo, Julia Kreutzer, Ahmet Üstün, Marzieh Fadaee, and Sara Hooker\.Aya dataset: An open\-access collection for multilingual instruction tuning\.In Lun\-Wei Ku, Andre Martins, and Vivek Srikumar \(eds\.\),*Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pp\. 11521–11567, Bangkok, Thailand, August 2024\. Association for Computational Linguistics\.doi:10\.18653/v1/2024\.acl\-long\.620\.URL[https://aclanthology\.org/2024\.acl\-long\.620/](https://aclanthology.org/2024.acl-long.620/)\.
- Soldaini et al\. \(2024\)Luca Soldaini, Rodney Kinney, Akshita Bhagia, Dustin Schwenk, David Atkinson, Russell Authur, Ben Bogin, Khyathi Chandu, Jennifer Dumas, Yanai Elazar, Valentin Hofmann, Ananya Jha, Sachin Kumar, Li Lucy, Xinxi Lyu, Nathan Lambert, Ian Magnusson, Jacob Morrison, Niklas Muennighoff, Aakanksha Naik, Crystal Nam, Matthew Peters, Abhilasha Ravichander, Kyle Richardson, Zejiang Shen, Emma Strubell, Nishant Subramani, Oyvind Tafjord, Evan Walsh, Luke Zettlemoyer, Noah Smith, Hannaneh Hajishirzi, Iz Beltagy, Dirk Groeneveld, Jesse Dodge, and Kyle Lo\.Dolma: an open corpus of three trillion tokens for language model pretraining research\.In Lun\-Wei Ku, Andre Martins, and Vivek Srikumar \(eds\.\),*Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pp\. 15725–15788, Bangkok, Thailand, August 2024\. Association for Computational Linguistics\.doi:10\.18653/v1/2024\.acl\-long\.840\.URL[https://aclanthology\.org/2024\.acl\-long\.840/](https://aclanthology.org/2024.acl-long.840/)\.
- Steingrímsson et al\. \(2023\)Steinþór Steingrímsson, Hrafn Loftsson, and Andy Way\.SentAlign: Accurate and Scalable Sentence Alignment\.In*Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations*, Singapore, Singapore, December 2023\. Association for Computational Linguistics\.
- Urieli et al\. \(2025\)Assaf Urieli, Amber Clooney, Michelle Sigiel, and Grisha Leyfer\.Jochre 3 and the yiddish ocr corpus\.*arXiv preprint arXiv:2501\.08442*, 2025\.
- U\.S\. Copyright Office \(2025\)U\.S\. Copyright Office\.Copyright and artificial intelligence: Part 3 – generative ai training\.Report, U\.S\. Copyright Office, May 2025\.URL[https://www\.copyright\.gov/ai/](https://www.copyright.gov/ai/)\.
- Webber et al\. \(2022\)Jacob Webber, Samuel K Lo, and Isaac L Bleaman\.Reyd\-the first yiddish text\-to\-speech dataset and system\.In*INTERSPEECH*, pp\. 2363–2367, 2022\.
- Wu et al\. \(2025\)Minghao Wu, Weixuan Wang, Sinuo Liu, Huifeng Yin, Xintong Wang, Yu Zhao, Chenyang Lyu, Longyue Wang, Weihua Luo, and Kaifu Zhang\.The bitter lesson learned from 2,000\+ multilingual benchmarks, 2025\.URL[https://arxiv\.org/abs/2504\.15521](https://arxiv.org/abs/2504.15521)\.
- Xue et al\. \(2021\)Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al\-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel\.mT5: A massively multilingual pre\-trained text\-to\-text transformer\.In Kristina Toutanova, Anna Rumshisky, Luke Zettlemoyer, Dilek Hakkani\-Tur, Iz Beltagy, Steven Bethard, Ryan Cotterell, Tanmoy Chakraborty, and Yichao Zhou \(eds\.\),*Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies*, pp\. 483–498, Online, June 2021\. Association for Computational Linguistics\.doi:10\.18653/v1/2021\.naacl\-main\.41\.URL[https://aclanthology\.org/2021\.naacl\-main\.41/](https://aclanthology.org/2021.naacl-main.41/)\.
- Yang et al\. \(2025\)An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al\.Qwen3 technical report\.*arXiv preprint arXiv:2505\.09388*, 2025\.
- Zhang et al\. \(2025\)Mike Zhang, Max Müller\-Eberstein, Elisa Bassignana, and Rob van der Goot\.SnakModel: Lessons learned from training an open Danish large language model\.In Richard Johansson and Sara Stymne \(eds\.\),*Proceedings of the Joint 25th Nordic Conference on Computational Linguistics and 11th Baltic Conference on Human Language Technologies \(NoDaLiDa/Baltic\-HLT 2025\)*, pp\. 812–825, Tallinn, Estonia, March 2025\. University of Tartu Library\.ISBN 978\-9908\-53\-109\-0\.URL[https://aclanthology\.org/2025\.nodalida\-1\.80/](https://aclanthology.org/2025.nodalida-1.80/)\.

## Appendix AKashes\-mt Sentence Alignment

![Refer to caption](https://arxiv.org/html/2608.05850v1/x1.png)\(a\)Forverts
![Refer to caption](https://arxiv.org/html/2608.05850v1/x2.png)\(b\)In geveb

Figure 2:The distribution of the portion of initially aligned sentences in all sentences in each document\. We selected documents with at least35%35\\%sentences aligned to filter out documents that could be incorrectly matched\.![Refer to caption](https://arxiv.org/html/2608.05850v1/x3.png)\(a\)Forverts
![Refer to caption](https://arxiv.org/html/2608.05850v1/x4.png)\(b\)In geveb

Figure 3:The distribution of SentAlign similarity score across initially aligned documents\. We selected a threshold of 0\.65 to filter out pairs that could be an inaccurate translation\.[Figure 2](https://arxiv.org/html/2608.05850#A1.F2)shows the distribution of the share of sentences that were initially aligned across all documents\. We filter out documents with less than35%35\\%sentences aligned\. In[Figure 3](https://arxiv.org/html/2608.05850#A1.F3), we present the distribution of similarity scores between pairs of sentences aligned with SentAlign\(Steingrímsson et al\.,[2023](https://arxiv.org/html/2608.05850#bib.bib33)\)\. We filter sentences with scores below0\.650\.65that make up the lower tail of the distribution\.

## Appendix BFew\-shot full result

Table 6:Results across all tasks and shot counts\.Bold= best per row\.
## Appendix CKashesExamples

Tables[7](https://arxiv.org/html/2608.05850#A3.T7)–[10](https://arxiv.org/html/2608.05850#A3.T10)contain examples from each dataset and task inKashes, presented in the format used during evaluation\. The quoted English text below each example is provided only as translation and explanation for the reader and is not used as part of the evaluation input\.

Table 7:Examples from the Aya Collection tasks \(Originally Machine\-translated to Yiddish\), formatted as they appear in the few\-shot prompts\.Table 8:Examples from the UD Yiddish\-YiTB treebank tasks, formatted as they appear in the few\-shot prompts\.Table 9:Examples from the three NER datasets, formatted as they appear in the few\-shot prompts\.Table 10:Examples from the machine translation task, formatted as they appear in the few\-shot prompts\.Example 1:English:The sons, the three thieves, loved their mother very much\.Yiddish:די זין, די דרײַ גנבֿים, האָבן זײער ליב געהאַט די מאַמע\.Hebrew\-origin “gnavim” \(thieves\), idiomatic “mame” \(mother\)—folktale narrative\.Example 2:English:“In that case,” said the head of the court, “Let’s ask at the prayer house where he used to pray\.”Yiddish:„אױב אַזױ, — האָט דער ראָש־בית־דין געזאָגט, — זאָל מען שיקן פֿרעגן אין בית־מדרש, װוּ אײַער מאַן פֿלעגט דאַװנען\.”English uses generic terms; Yiddish uses Hebrew\-Aramaic loanwords: “rosh\-beys\-din”, “beys\-medresh”, “davnen”\.
## Appendix DAuditing the Yiddish Split of mC4

This appendix details the analysis summarized in[Section 3\.1](https://arxiv.org/html/2608.05850#S3.SS1)\. The split contains 143,708 pages from 7,621 distinct domains, each labelledyiby the CLD3 identifier used in mC4’s construction\(Xue et al\.,[2021](https://arxiv.org/html/2608.05850#bib.bib38)\)\. We audit all 143,708 pages for two failure modes, Hebrew misidentified as Yiddish through the shared script, and machine translation;[Table 11](https://arxiv.org/html/2608.05850#A4.T11)gives the resulting composition\.

Hebrew and Yiddish share a script, and general\-purpose language identifiers are unreliable on the pair\. We therefore trained a dedicated Hebrew/Yiddish classifier: a character n\-gram Multinomial Naive Bayes model \(n = 2–4\) trained on 108,000 sentences per class from Yiddish and Hebrew Wikipedia, with input normalized to Hebrew\-script characters and diacritic\-stripped copies added to training\. On 12,000 held\-out sentences it reaches macro\-F1 0\.9968 \(95% CI\[0\.9957,0\.9979\]\[0\.9957,0\.9979\]\) and stays above 0\.995 without diacritics and on 50\-character snippets\.

Because no gold labels exist for the crawl, we validated it against silver labels derived from URL locale structure \(yi\./he\./iw\.subdomains and paths\), a signal it does not observe, available for 50,030 pages: it agrees on 99\.6% of them\. Applied to the full split, the classifier marks 31,485 pages \(21\.9%\) as Hebrew rather than Yiddish \([Table 11](https://arxiv.org/html/2608.05850#A4.T11)\)\.

The second failure mode is machine translation\. Machine\-translated web content has a characteristic URL fingerprint: the same site served under many language codes, with the Yiddish “edition” appearing as ayi\.oryid\.subdomain, a language\-code path segment \(/yi/,LANG\-yi\), or a language\-code query parameter \(e\.g\.,yi\.itsmygame\.org,yid\.feminineclub\.com\)\. Pages were assigned to*known Yiddish sources*\(manually verified native outlets such asivelt\.com,kaveshtiebel\.com, andyiddish\.forward\.com\) or to*suspected machine translation*\(the URL fingerprint above\)\. We manually inspected several hundred domains, ranked by page count, and read sampled pages from each category; the suspected\-MT domains were predominantly spam\-like sites \(gambling, gaming, explicit\-content, and content\-farm pages rendered into dozens of languages\)\.

The URL fingerprint alone cannot separate a genuine bilingual outlet from an MT template site;yiddish\.forward\.comcarries a language\-labeled subdomain just asyi\.itsmygame\.orgdoes\. We therefore backed it with a*locale\-sibling count*: for each domain, we queried a Common Crawl index contemporary with the corpus’s construction and counted the distinct language editions \(locale\-code subdomains or path prefixes\) under which the same site appears\. An MT template site serves the same content under many language codes, while a native outlet maintains at most one or two editions\. On samples of 50 domains from the suspected\-MT and native categories, MT\-flagged domains appeared under a mean of 25 language editions, against fewer than one for native domains\. Genuinely bilingual publishers such asyiddish\.forward\.comare thus cleared by measurement rather than by prior knowledge of the Yiddish web\.

The locale\-sibling count can still produce false positives, since some institutions legitimately publish in many languages; governmental portals are a typical example, and some carry genuine Yiddish pages where Yiddish is a recognized minority language\. Manual investigation, however, showed that from the sample almost all domains confirmed by both the fingerprint and the sibling count came from contexts almost surely unrelated to any Yiddish\-speaking community, such as gambling, gaming, adult\-content, and commercial content\-farm sites\. While this assessment is somewhat subjective, it is unusually dependable for Yiddish: the language’s present\-day native speakers are concentrated in Hasidic and other Haredi communities, and content of this character is highly unlikely to be produced by or addressed to that community publicly\.

Beyond the pages captured by the native whitelist, the MT fingerprint, and the Hebrew classifier lies a long tail of 8,871 pages \(6\.2% of the split\) across some 1,600 domains: very short Hebrew\-script fragments that resist reliable identification, multilingual template pages, and small uncatalogued native Yiddish sources\. Both the native and the MT shares of[Table 11](https://arxiv.org/html/2608.05850#A4.T11)are therefore lower bounds\.

Table 11:Composition of the mC4 Yiddish split, all 143,708 pages\. The first two categories are assigned by source URL; the remaining pages are split by the Hebrew/Yiddish classifier\. LK rate is the mean per\-document share of loshn\-koydesh \(Hebrew/Aramaic\-origin\) content words \(±\\pmSE\); it is not reported for the Hebrew category, where Hebrew text produces spuriously high matches\.More than one page in five in the mC4 Yiddish split is Hebrew, not Yiddish, and well under half comes from native Yiddish sources\.

Finally, the loshn\-koydesh \(LK\) rate serves as one more independent tool \(the lexicon and matching procedure are described in[AppendixE\.1](https://arxiv.org/html/2608.05850#A5.SS1)\)\. Suspected\-MT pages average an LK rate of 3\.6%, roughly a third of the 10\.2% of native sources, consistent with translationese that substitutes Germanic or internationalist equivalents for Hebrew/Aramaic\-origin vocabulary\. No rate is reported for the Hebrew category, where Hebrew text trivially matches the Hebrew\-origin lexicon\. The residual category shows the mixture its composition implies, with a median LK of 5\.7% against a mean of 9\.4%\.

## Appendix EAnalysis Methodology

This appendix provides full technical details for the linguistic competence analyses presented in[Section 8](https://arxiv.org/html/2608.05850#S8)\. All analyses draw on two primary resources: a loshn\-koydesh \(LK\) identification lexicon and morphological annotations from the UD Yiddish\-YITB treebank\.

### E\.1Loshn\-Koydesh Identification

#### Lexicon\.

We identify Hebrew/Aramaic\-origin vocabulary using the Bleaman/Niborski loshn\-koydesh pronunciation lexicon,999[https://github\.com/ibleaman/loshn\-koydesh\-pronunciation](https://github.com/ibleaman/loshn-koydesh-pronunciation)a digitized version of Eliezer Niborski’s reference work\. The lexicon maps Hebrew orthographic forms to their Yiddish phonetic transcriptions\. We extract 5,437 single\-word entries and 2,926 compound phrases \(entries containing the Hebrew maqaf ־ or a hyphen delimiter\)\.

#### Tokenization\.

Yiddish text is tokenized into words using a regular expression over Unicode Hebrew\-script code points: characters in the ranges U\+05D0–U\+05EA \(Hebrew letters\), U\+05F0–U\+05F4 \(Yiddish ligatures\), U\+FB1D–U\+FB4E \(Hebrew presentation forms\), and U\+05B0–U\+05C7 \(Hebrew diacritical marks/nikud\)\. Each maximal contiguous span of such characters constitutes a token\.

#### Content word filtering\.

Tokens are classified as content words if they satisfy two conditions: \(1\) the token contains≥\\geq3 characters after removal of all combining diacritical marks \(Unicode category Mn\), and \(2\) the diacritic\-stripped form is not in a stopword list of common Germanic\-origin function words\. The stopword list contains the following items:

> די, א, אין, צו, פאר, אז, ס, זיין, דאך, צום, פ, מיט, פון, איז, האט, ניט, נישט, אויך, שוין, נאר, ווי, וואס, דאס, דאָס, איך, ער, זי, מיר, איר, דער, דעם, עס, מען, זײ, אים, זיך, אַ, אַן, און, אָדער, אָבער, דאָ, נאָך, פֿון, פֿאַר

#### LK matching\.

Each content word is tested against the lexicon using a two\-pass procedure:

1. 1\.NFC\-normalized exact match: the token is Unicode NFC\-normalized and compared to the set of NFC\-normalized lexicon entries\.
2. 2\.Diacritic\-stripped fallback: if no NFC match is found, all combining marks \(Unicode category Mn\) are removed from both the token and lexicon entries, and comparison is repeated\. This handles variation innikud\(vowel pointing\) between sources\.

For compound phrases, a sliding window of sizenn\(for eachnn\-gram size present in the lexicon\) is passed over the full token stream\. If annn\-gram matches a compound entry \(by NFC or stripped comparison\), it is recorded as a single LK item\. Individual components of compound entries are*not*treated as standalone LK words, preventing false positives from Germanic words that happen to appear within LK compounds\.

### E\.2LK Vocabulary Production in Translation

This analysis measures the rate at which models produce loshn\-koydesh vocabulary in English→\\rightarrowYiddish translation, corresponding to the LK production results in[Section 8](https://arxiv.org/html/2608.05850#S8)\.

#### Data\.

We use 5,287 English–Yiddish sentence pairs from theKashes\-mt dataset, drawn from the In Geveb \(4,292\) and Forward \(995\) sources\. Each sentence pair is translated by two models \(MameLoshnLMand Llama 3\.1 8B\) at a 5\-shot setup\.

#### Metrics\.

For each generated translation, we apply the tokenization, content word filtering, and LK matching pipeline described in[AppendixE\.1](https://arxiv.org/html/2608.05850#A5.SS1)\. We then compute three metrics:

- •LK content word rate: the proportion of content words identified as LK in the generated text\. Computed per sentence and reported as the mean±\\pmstandard error across alln=5,287n=5\{,\}287sentences: LK rate=1n​∑i=1n\|\{LK tokens in sentence​i\}\|\|\{content tokens in sentence​i\}\|\\text\{LK rate\}=\\frac\{1\}\{n\}\\sum\_\{i=1\}^\{n\}\\frac\{\|\\\{\\text\{LK tokens in sentence \}i\\\}\|\}\{\|\\\{\\text\{content tokens in sentence \}i\\\}\|\}
- •LK sentence match rate: restricted to thenLK=3,112n\_\{\\text\{LK\}\}=3\{,\}112sentences where the gold reference contains≥\\geq1 LK content word\. For each such sentence, we check whether the model’s translation contains at least one LK word that also appears in the gold \(after diacritic\-stripped comparison\)\. Reported as: Match rate=\|\{i:model LKi∩gold LKi≠∅\}\|nLK\\text\{Match rate\}=\\frac\{\|\\\{i:\\text\{model LK\}\_\{i\}\\cap\\text\{gold LK\}\_\{i\}\\neq\\emptyset\\\}\|\}\{n\_\{\\text\{LK\}\}\}
- •Per\-word LK recall: for each gold\-LK sentence, the fraction of unique gold LK word types reproduced by the model \(after diacritic stripping\)\. Reported as the mean±\\pmSE across gold\-LK sentences: Recall=1nLK​∑i=1nLK\|model LKi∩gold LKi\|\|gold LKi\|\\text\{Recall\}=\\frac\{1\}\{n\_\{\\text\{LK\}\}\}\\sum\_\{i=1\}^\{n\_\{\\text\{LK\}\}\}\\frac\{\|\\text\{model LK\}\_\{i\}\\cap\\text\{gold LK\}\_\{i\}\|\}\{\|\\text\{gold LK\}\_\{i\}\|\}

#### LK→\\rightarrowGermanic substitution analysis\.

To characterize what models produce instead of LK words, we examined all 72 LK words with gold frequency≥20\\geq 20at 5\-shot\. For each LK word, we collected all sentences where \(a\) the gold contains the LK word but \(b\) the model’s output does not\. From these sentences, we extracted the model’s non\-gold content words as candidate substitutions\. Germanic equivalents were identified*empirically from model output*—not pre\-specified—and then manually verified as genuine synonym pairs\.

#### Statistical tests\.

We compareMameLoshnLMand Llama 3\.1 8B using two paired tests:

- •LK rate: pairedtt\-test on per\-sentence LK rates \(same sentence, two models\)\.
- •Match rate: McNemar’s test on sentence\-level binary outcomes \(bb=MameLoshnLMmatches but Llama does not;cc= Llama matches butMameLoshnLMdoes not\)\. Implemented as a two\-sided exact binomial test:p=binomtest​\(b,b\+c,0\.5\)p=\\texttt\{binomtest\}\(b,\\,b\+c,\\,0\.5\)\.
- •Recall: pairedtt\-test on per\-sentence recall values\.

### E\.3Morphological Analysis via Lemmatization

This analysis uses lemmatization as a controlled probe of morphological knowledge\. The task is to map each inflected Yiddish word to its citation form \(lemma\), where each token presents a single morphological operation with an unambiguous gold answer\. This corresponds to the morphological results in[Section 8](https://arxiv.org/html/2608.05850#S8)\.

#### Data\.

The test set consists of 929 sentences \(7,499 tokens\) from the UD Yiddish\-YITB treebank\.101010[https://github\.com/UniversalDependencies/UD\_Yiddish\-YITB](https://github.com/UniversalDependencies/UD_Yiddish-YITB)Gold annotations include surface form, lemma, universal POS tag \(UPOS\), and dependency relations in CoNLL\-U format\. Each model generates lemmatization predictions in a structured word→\\rightarrowlemma format at 5\-shot counts\.

#### Prediction postprocessing\.

Model outputs are postprocessed before evaluation: \(1\) trailing punctuation characters \(\.,;:\!?’̈\)\]…\) are stripped from predicted lemmas using a character\-level regex; \(2\) indefinite articles \(אַ/אַן\) preceding a word are detached and evaluated as separate pairs\.

#### Token classification\.

Every token in the test set is categorized along two axes:

1. 1\.Etymology: LK vs\. non\-LK\. A token is classified as LK if either its surface form or its gold lemma matches the Bleaman/Niborski lexicon \(after NFC normalization and diacritic\-stripped fallback\), and neither form is in the stopword list\. This dual check ensures that tokens whose lemma is LK but whose surface form has diverged \(or vice versa\) are captured\.
2. 2\.Transformation type:*change*tokens \(surface≠\\neqlemma\) vs\.*identity*tokens \(surface==lemma\)\. Change tokens are the informative subset, since identity tokens can be trivially handled by copying the input\.

Dataset composition: 565 LK tokens \(7\.5%\) and 6,934 non\-LK tokens \(92\.5%\)\. Among LK tokens, 48\.1% require a lemma change vs\. 33\.5% of non\-LK tokens\.

#### LK lemmatization metrics\.

We reportchange accuracyseparately for LK and non\-LK strata:

ChangeAccstratum=\|\{t∈stratum:tword≠tlemma∧tpred=tlemma\}\|\|\{t∈stratum:tword≠tlemma\}\|\\text\{ChangeAcc\}\_\{\\text\{stratum\}\}=\\frac\{\|\\\{t\\in\\text\{stratum\}:t\_\{\\text\{word\}\}\\neq t\_\{\\text\{lemma\}\}\\wedge t\_\{\\text\{pred\}\}=t\_\{\\text\{lemma\}\}\\\}\|\}\{\|\\\{t\\in\\text\{stratum\}:t\_\{\\text\{word\}\}\\neq t\_\{\\text\{lemma\}\}\\\}\|\}The LK–non\-LK gap is the arithmetic difference between the two strata’s change accuracies\.

### E\.4Morphological Category Breakdown

To understand*which*morphological phenomena drive the overall gap, we classify all 2,593 change tokens into six categories using the gold CoNLL\-U UPOS tags for disambiguation:

1. 1\.ge\-participle\(n=236n=236\): the surface form begins with גע \(or a verbal prefix followed by גע, e\.g\., אָפּגע, אױסגע, אַרײַנגע\), the lemma does not contain גע, and the UPOS is VERB or AUX\. We recognize 23 prefixedge\-patterns \(e\.g\., אָפּגע, אױסגע, אַרײַנגע, אַרױסגע, אונטערגע, צוגע, among others\)\.
2. 2\.Adjective declension\(n=173n=173\): the surface form ends in ער, ע, or ן; stripping that suffix yields the lemma; and UPOS is ADJ\.
3. 3\.Hebrew\-origin plural\(n=53n=53\): the surface form ends in ים or ות \(Hebrew masculine/feminine plural suffixes\) while the lemma does not, and UPOS is NOUN or ADJ\.
4. 4\.Determiner paradigm\(n=291n=291\): the gold lemma is דער and the surface form is one of the declined forms די, דאָס, or דעם\.
5. 5\.Verb conjugation\(n=1,084n=1\{,\}084\): the lemma ends in ן or ען, UPOS is VERB or AUX, and the token does not fall into thege\-participle category\. Both surface form and lemma must be≥\\geq3 characters\.
6. 6\.Other\(n=793n=793\): all remaining change tokens \(noun plurals with Germanic suffixes, pronoun inflections, spelling variants, etc\.\)\.

Categories are tested in priority order \(1–6\), so each token is assigned to exactly one category\. Change accuracy is computed per category as defined in[AppendixE\.3](https://arxiv.org/html/2608.05850#A5.SS3)\.

### E\.5Determiner Case System

Yiddish has a three\-gender, four\-case definite article system in which all declined forms share the citation lemma דער\. We extract all tokens from the lemmatization test set where the gold lemma is דער, grouped by surface form:

Lemmatization accuracy is computed per form\. For statistical testing, we exclude the identity case \(דער→\\rightarrowדער\) and test each declined form individually and aggregated across all three declined forms using McNemar’s exact test\.

### E\.6Auxiliary Verb Paradigms

We extract all tokens tagged as AUX in the gold CoNLL\-U annotations \(1,000 tokens total\) and group them by gold lemma into three paradigms:

The key comparison is between זײַן \(suppletive: inflected forms bear no surface resemblance to the lemma\) and האָבן \(regular: האָט→\\rightarrowהאָבן is a straightforward suffix change\)\. This serves as a within\-task control: if a model’s advantage is specific to language\-specific knowledge rather than general lemmatization ability, it should appear on the suppletive paradigm but not the regular one\.

### E\.7Statistical Testing

All pairwise model comparisons \(MameLoshnLMvs\. Llama 3\.1 8B\) on lemmatization useMcNemar’s exact test\(two\-sided\)\. The test operates on paired binary outcomes: for each token \(or sentence, in the translation analysis\), we record whether each model’s prediction is correct\. We count:

- •bb: tokens whereMameLoshnLMis correct and Llama is incorrect \(discordant, favoringMameLoshnLM\)
- •cc: tokens where Llama is correct andMameLoshnLMis incorrect \(discordant, favoring Llama\)

Concordant pairs \(both correct or both incorrect\) are uninformative and excluded\. Under the null hypothesis that both models are equally likely to be correct on discordant pairs,b∼Binomial​\(b\+c,0\.5\)b\\sim\\text\{Binomial\}\(b\+c,\\,0\.5\)\. The two\-sidedpp\-value is computed viascipy\.stats\.binomtest\(b, b\+c, 0\.5\)\. We report significance at three levels:∗\{\*\}\(p<0\.05p<0\.05\),∗⁣∗\{\*\*\}\(p<0\.01p<0\.01\),∗⁣∗⁣∗\{\*\*\*\}\(p<0\.001p<0\.001\)\.

McNemar’s test is appropriate here because observations are paired \(both models predict on the same token\) and outcomes are binary \(correct/incorrect\)\. It is more powerful than unpaired tests because it controls for item difficulty\.

For the translation analysis, we additionally usepairedtt\-testson continuous per\-sentence metrics \(LK rate, LK recall\), since these are real\-valued rather than binary\.

### E\.8Summary of All Metrics

[Table 12](https://arxiv.org/html/2608.05850#A5.T12)consolidates all metrics from the analyses above, comparingMameLoshnLMand Llama 3\.1 8B across both the LK translation analysis \(5\-shot\) and the morphological analysis \(5\-shot lemmatization\)\.

Table 12:Complete comparison ofMameLoshnLMvs\. Llama 3\.1 8B across all analysis dimensions\. Translation metrics are at 5\-shot \(n=5,287n\{=\}5\{,\}287sentence pairs; match and recall computed overn=3,112n\{=\}3\{,\}112gold\-LK sentences\)\. Morphological metrics are at 5\-shot on the UD Yiddish\-YITB test set\. McNemarbb:ccdenotes discordant pairs favoringMameLoshnLMvs\. Llama\. Allpp\-values are two\-sided\.DomainMetricMameLoshnLMLlama 3\.1GoldTest statisticpp\-valueLK vocabulary production \(eng→\\toyid translation, 5\-shot\)LK content word rate \(%\)4\.71\.66\.2t=28\.0t\{=\}28\.0<10−161<10^\{\-161\}LK sentence match rate \(%\)52\.416\.0—bb:cc= 1181:47<10−229<10^\{\-229\}Per\-word LK recall \(%\)35\.69\.3—t=36\.8t\{=\}36\.8<10−246<10^\{\-246\}Morphological analysis \(lemmatization, 5\-shot\)ge\-participles \(n=236n\{=\}236\)50\.85\.1—bb:cc= 110:2<10−29<10^\{\-29\}Hebrew\-origin plurals \(n=53n\{=\}53\)30\.22\.3—bb:cc= 12:0<0\.001<0\.001Adj\. declension \(n=173n\{=\}173\)68\.541\.8—bb:cc= 42:3<10−9<10^\{\-9\}Determiners \(n=291n\{=\}291\)28\.215\.5—bb:cc= 55:18<10−5<10^\{\-5\}Verb conjugation \(n=1,084n\{=\}1\{,\}084\)46\.733\.8—bb:cc= 196:56<10−18<10^\{\-18\}Auxiliary verb paradigms \(lemmatization, 5\-shot\)זײַן — suppletive \(n=460n\{=\}460\)20\.08\.9—bb:cc= 63:12<10−9<10^\{\-9\}האָבן — regular \(n=163n\{=\}163\)84\.784\.7—bb:cc= 14:141\.01\.0\(n\.s\.\)
### E\.9Most Frequent Loshn\-Koydesh Words in Gold References

[Table 13](https://arxiv.org/html/2608.05850#A5.T13)lists the 15 most frequent LK content words in the gold Yiddish references, along with the absolute number of times each word appears in the 5\-shot translations ofMameLoshnLMand Llama 3\.1 8B \(across alln=5,287n\{=\}5\{,\}287sentence pairs from the In Geveb and Forward sources\)\.

Table 13:Top 15 LK words by gold frequency \(absolute counts across 5,287 5\-shot translations\)\. Common Yiddish vocabulary items \(e\.g\.,efsher,ponim,kedey\) are largely absent from Llama’s output, while proper nouns and cultural terms \(e\.g\.,rebe,khane,yisroel\) appear at comparable rates\.

## Appendix FEffect of related\-language mixing

Table 14:Results overKashesofMameLoshnLMand Llama 3\.1 compared to models trained on data mixtures with more Yiddish\-related languages\. All results are in a 5\-shot setting, scaled 0–100\.Bold= best per row\.Table 15:Training mixture composition across the three continued\-pretraining settings\. Percentages are shown over words and over training tokens\.