@MaximeRivest: The recipe is simple: 1. Find repeatable tasks. 2. Use whatever model/size you need to get the task done satisfactorily…
Summary
推文分享了一套实用配方:找到可重复任务,用任意模型完成并记录运行结果,然后微调最小的开源权重模型(如 FastConformer)达到目标精度,从而以极小算力获得接近 SOTA 的速度与性能。作者以自己用 Parakeet 采集数据训练 FastConformer 的实际经验为例。
View Cached Full Text
Cached at: 10/19/26, 01:24 PM
The recipe is simple:
- Find repeatable tasks.
- Use whatever model/size you need to get the task done satisfactorily.
- Log the runs.
- Fine-tune the smallest open-weight model that gets you the accuracy you want.
- Enjoy the benefits of SOTA-level performance at blazing speed on tiny compute.
In that specific case below, I used Parakeet for several months, saved all my audio and transcriptions, and trained the tiny FastConformer on it.
Similar Articles
@MaximeRivest: This is a really good question. I am not sure what to recommend, to be honest. If I tried to boil it down, it would be:…
Maxime Rivest回复网友关于如何入门模型微调的提问,给出了从简单分类任务起步、逐步构建自定义任务、利用coding agent辅助实验的五步学习路径建议。
@MaximeRivest: wow, can't wait to try it. we can do so much in specialized 8b!
Ai2(@allen_ai)发布开源模型 AstaBrief 8B,可将复杂研究问题与文献片段转化为带引用的报告,支持本地运行,开放权重与训练数据;@MaximeRivest 表示期待尝试小型专用 8B 模型的潜力。
@vintcessun: Pretraining can be this cost-effective? Train a usable 1B base model from scratch for ~$1000, slashing compute and data by hundreds of times. The key isn't brute-force compute, but hierarchical recursive architecture plus latent space reasoning, combined with PrefixLM packing and FA3 to maximize efficiency. Sounds insane, but the paper and code are open-sourced.
HRM-Text released a 1B-parameter base model, claiming it can be pretrained from scratch for only ~$1000, reducing compute and data volume by hundreds of times. It employs efficient techniques such as hierarchical recursive architecture, latent space reasoning, and PrefixLM packing. The paper and code are open-sourced.
@HuggingPapers: Recursive Self-Rewrite scales model capability on complex tasks A single base model (Qwen-3.8-27B) discovers solutions …
HuggingPapers 介绍了一篇关于 Recursive Self-Rewrite 的研究:单一基础模型 Qwen-3.8-27B 在多种 harness 下发现解法并将其改写为可复用的训练轨迹,使 Terminal-Bench 2 的 Pass@3 从 57.0% 提升到 74.2%。
@BohuTANG: I used to try cross-model mutual review, but that was too slow for me. Now I've discovered a new method: /harden, which achieves great results with two rounds of convergence on the same model. Interested folks can try this skill.
BohuTANG introduces /harden, a method for same-model two-round convergence, and highlights the evot agent engine which completes complex tasks with fewer tokens and lower cost than alternatives like Claude Code.