xiaoyuchen1/Qwen3-8B-GRPO-MATH

TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 26, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The xiaoyuchen1/Qwen3-8B-GRPO-MATH model is an 8 billion parameter Qwen3-8B base model fine-tuned with Group-Relative Advantage (GRPO) on the Hendrycks MATH training split, utilizing verifiable final-answer rewards. This research checkpoint, developed by xiaoyuchen1, is specifically optimized for mathematical reasoning tasks, demonstrating improved performance on benchmarks like AIME and MATH-500 compared to its base model. It is intended for research into RL-for-reasoning and on-policy distillation, rather than deployment.

Loading preview...

Overview

This model, xiaoyuchen1/Qwen3-8B-GRPO-MATH, is an 8 billion parameter variant of the Qwen3-8B base model. It has been fine-tuned using the Group-Relative Advantage (GRPO) algorithm on the Hendrycks MATH training dataset, specifically leveraging verifiable final-answer rewards. Developed as a research checkpoint, its primary purpose is to explore the effectiveness of RL gains in on-policy distillation for smaller student models.

Key Capabilities & Performance

  • Enhanced Mathematical Reasoning: The model shows significant improvements in mathematical problem-solving. On the AIME 2024 + 2025 benchmark, it achieved an avg@16 score of 34.27 and pass@16 of 66.7, substantially outperforming the base Qwen3-8B (22.09 avg@16, 46.7 pass@16).
  • MATH-500 Performance: It also improved on the MATH-500 benchmark, scoring 85.50 avg@16 and 94.6 pass@16 compared to the base model's 80.84 avg@16 and 92.8 pass@16.
  • Reward Mechanism: Training utilized a rule-based final-answer equivalence reward (--rm-type math), which was verified to be highly accurate for the MATH dataset's diverse answer formats, including symbolic LaTeX.

Limitations & Considerations

  • Research Checkpoint: This is a short-budget research checkpoint (120 optimizer updates, 0.26 epoch) and not a fully converged or tuned release.
  • Length Inflation: The model tends to generate longer responses (mean length 5,203 tokens vs. 2,325 for base), which can lead to truncation and zero scores if the context budget is too small. Optimal performance on AIME requires a 16,384-token budget.
  • Intended Use: It is explicitly not intended for deployment but rather for research in RL-for-reasoning and on-policy distillation.

Finetunes
1 models