CHIMERA: Compact Synthetic Data for Generalizable LLM Reasoning

Researchers introduce CHIMERA, a compact 9K-sample synthetic dataset that enables smaller AI models to achieve reasoning performance comparable to much larger models. The dataset addresses key challenges in training reasoning-capable LLMs through automated generation and cross-validation across 8 scientific disciplines.

• →CHIMERA dataset contains only 9K samples but enables a 4B parameter model to match performance of 235B parameter models on reasoning benchmarks.
• →The dataset spans 8 major scientific disciplines with over 1K fine-grained topics, addressing limited domain coverage in existing reasoning datasets.
• →Uses fully automated evaluation pipeline with strong reasoning models for cross-validation, solving expensive human annotation bottleneck.
• →Demonstrates that compact, high-quality synthetic data can be more effective than massive datasets for training reasoning capabilities.
• →The resulting model shows strong performance on challenging benchmarks including GPQA-Diamond and AIME competitions.