Research map / Large language models / Agents, code and math

DataGen: Unified Synthetic Dataset Generation via Large Language Models

ICLR 2025 · Large language models › Agents, code and math · approach: Code generation benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DataGen, an LLM-powered framework that generates controllable, diverse, and truthful text datasets using attribute-guided generation, group checking, code-based label verification, and RAG-based factual validation, and applies it to LLM benchmarking and data augmentation.

Problem

Existing LLM-driven dataset generation struggles with generalization, controllability, diversity, and truthfulness (hallucinated or incorrect labels), and often only perturbs existing items or is limited to specific formats.

Approach

A unified pipeline that takes a base dataset, a natural-language description, and optional user constraints, then combines few-shot clustering-based example selection, attribute-guided generation, embedding-based group checking, self-reflection/enhancement, code-verified math labels, and Wikipedia RAG validation, plus difficulty enhancement as post-processing.

Key results

• Code-based math verification raised label correctness on generated GSM8K from 44% to 88% (reported as 92% in text), and RAG validation corrected 4.2% of examples.
• Attribute-guided generation and group checking raised remote-clique diversity from 0.695 to 0.743 on GPT-4-generated data.
• Data augmentation with 200 DataGen samples improved performance on all ten evaluated datasets, and the genset alpaca-tuned Llama3 model scored 6.403 on MT-Bench versus 5.259 for the original-alpaca model.

Measured effects reported in the paper

Surprising findings

• Most evaluated LLMs score lower on GPT-4-generated GSM8K than on the original GSM8K, suggesting overstated reasoning ability.
• Llama2-7b's TruthfulQA performance declined after fine-tuning on generated data.

Limitations

• Error analysis shows factuality errors (e.g., 41% on GSM8K, 79% on TruthfulQA) and question-quality errors remain in generated data.
• Limited gains on knowledge-intensive datasets such as MMLU and TruthfulQA after augmentation.
• Generation cost up to about $0.20 per item with RAG and code validation, and self-alignment or weak-to-strong settings were not studied.

Experimental setup

Datasets: GSM8K MMLU TruthfulQA HellaSwag MetaTool MultiNLI ARC-Challenge BoolQ BBH alpaca

Benchmarks: GSM8K MMLU TruthfulQA HellaSwag MetaTool ARC-Challenge BoolQ BBH MT-Bench

Metrics: accuracy remote-clique diversity self-BLEU entity overlap rate error proportion

Baselines: DyVal 2 DyVal S3Eval MetaMath Qameleon

Builds on: DyVal 2 MetaMath LLaMA-Factory

Closest papers by meaning

• Towards a Theoretical Understanding of Synthetic Data in LLM Post-Training: A Reverse-Bottleneck Perspective (ICLR 2025 · similarity 0.88)
• Controlled Text Generation via Language Model Arithmetic (ICLR 2024 · similarity 0.87)
• It's All Training: A Fully Synthetic Single-Stage Recipe for LLMs (NeurIPS 2026 · similarity 0.87)
• Automating Dataset Updates Towards Reliable and Timely Evaluation of Large Language Models (NeurIPS 2024 · similarity 0.87)
• ORACLE: Optimizing Reasoning Abilities of Large Language Models via Constraint-Led Synthetic Data Elicitation (AAAI 2026 · similarity 0.87)
• Not All LLM-Generated Data Are Equal: Rethinking Data Weighting in Text Classification (ICLR 2025 · similarity 0.86)