GraphGen: Enhancing Supervised Fine-Tuning for LLMs with Knowledge-Driven Synthetic Data Generation

Papers with Code Trending Papers
knowledge-graph synthetic-data supervised-fine-tuning llm data-generation qa

Summary

GraphGen is a knowledge-graph-guided framework for generating synthetic QA data to improve supervised fine-tuning of LLMs, targeting knowledge gaps with multi-hop sampling and style-controlled generation. Experiments show it outperforms conventional synthetic data methods.

Fine-tuning for large language models (LLMs) typically requires substantial amounts of high-quality supervised data, which is both costly and labor-intensive to acquire. While synthetic data generation has emerged as a promising solution, existing approaches frequently suffer from factual inaccuracies, insufficient long-tail coverage, simplistic knowledge structures, and homogenized outputs. To address these challenges, we introduce GraphGen, a knowledge graph-guided framework designed for three key question-answering (QA) scenarios: atomic QA, aggregated QA, and multi-hop QA. It begins by constructing a fine-grained knowledge graph from the source text. It then identifies knowledge gaps in LLMs using the expected calibration error metric, prioritizing the generation of QA pairs that target high-value, long-tail knowledge. Furthermore, GraphGen incorporates multi-hop neighborhood sampling to capture complex relational information and employs style-controlled generation to diversify the resulting QA data. Experimental results on knowledge-intensive tasks under closed-book settings demonstrate that GraphGen outperforms conventional synthetic data methods, offering a more reliable and comprehensive solution to the data scarcity challenge in supervised fine-tuning. The code and data are publicly available at https://github.com/open-sciencelab/GraphGen.
Original Article
View Cached Full Text

Cached at: 08/03/26, 01:29 PM

Paper page - GraphGen: Enhancing Supervised Fine-Tuning for LLMs with Knowledge-Driven Synthetic Data Generation

Source: https://huggingface.co/papers/2505.20416 Published on May 26, 2025

Abstract

GraphGen is a knowledge graph-guided framework that addresses challenges in synthetic data generation for LLMs by constructing fine-grained knowledge graphs, targeting high-value knowledge gaps, and employing multi-hop sampling and style-controlled generation.

Fine-tuningforlarge language models(LLMs) typically requires substantial amounts of high-quality supervised data, which is both costly and labor-intensive to acquire. Whilesynthetic data generationhas emerged as a promising solution, existing approaches frequently suffer from factual inaccuracies, insufficient long-tail coverage, simplistic knowledge structures, and homogenized outputs. To address these challenges, we introduce GraphGen, aknowledge graph-guided framework designed for three key question-answering (QA) scenarios: atomic QA, aggregated QA, andmulti-hop QA. It begins by constructing a fine-grainedknowledge graphfrom the source text. It then identifiesknowledge gapsin LLMs using theexpected calibration errormetric, prioritizing the generation of QA pairs that target high-value, long-tail knowledge. Furthermore, GraphGen incorporatesmulti-hop neighborhood samplingto capture complex relational information and employs style-controlled generation to diversify the resulting QA data. Experimental results on knowledge-intensive tasks underclosed-book settingsdemonstrate that GraphGen outperforms conventional synthetic data methods, offering a more reliable and comprehensive solution to the data scarcity challenge in supervisedfine-tuning. The code and data are publicly available at https://github.com/open-sciencelab/GraphGen.

View arXiv pageView PDFProject pageGitHub1.17kAdd to collection

Get this paper in your agent:

hf papers read 2505\.20416

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2505.20416 in a model README.md to link it from this page.

Datasets citing this paper1

#### chenzihong/GraphGen-Data Preview• UpdatedJun 16, 2025 • 27 • 2

Spaces citing this paper1

Collections including this paper1

Similar Articles

Goal-Conditioned Supervised Learning for LLM Fine-Tuning

arXiv cs.LG

This paper proposes goal-conditioned supervised learning (GCSL) as an offline fine-tuning framework for LLMs, which treats feedback as an explicit goal and trains models via supervised learning with a novel goal formulation and natural-language goal representations. Evaluated on non-toxic generation, code generation, and recommendation, it outperforms standard offline baselines.