Linyuana/qwen3-0.6b-grpo-math-reasoning

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 11, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Linyuana/qwen3-0.6b-grpo-math-reasoning is a 0.8 billion parameter Qwen3-based language model, fine-tuned using Group Relative Policy Optimization (GRPO) on GSM8K and MATH datasets. This model is specifically optimized for mathematical reasoning tasks, demonstrating significant performance improvements over its SFT-only counterpart. It excels at generating verifiable step-by-step mathematical solutions, making it suitable for applications requiring accurate numerical problem-solving.

Loading preview...

Model Overview

Linyuana/qwen3-0.6b-grpo-math-reasoning is a 0.8 billion parameter Qwen3-based model specifically fine-tuned for mathematical reasoning. It leverages a novel training approach combining Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO), using a rule-based reward system without a separate reward model. This model is part of a reproduction study comparing GRPO, PPO, and DPO, highlighting GRPO's stability and performance advantages in this domain.

Key Capabilities

  • Enhanced Mathematical Reasoning: Achieves significantly higher pass rates on GSM8K and MATH benchmarks compared to base and SFT-only models.
  • GRPO Optimization: Utilizes a critic-free GRPO algorithm, which demonstrated superior stability and performance over PPO and DPO for this task.
  • Verifiable Outputs: Designed to produce solutions in a <think>...</think><answer>...</answer> format, facilitating extraction and verification of final answers.
  • Efficient Scale: Provides strong mathematical reasoning capabilities at a compact 0.8 billion parameter size.

Performance Highlights

On held-out samples, the model achieved:

  • GSM8K: 67.7% pass@1 and 85.0% pass@8 (compared to 38.9% and 78.4% for SFT only).
  • MATH: 48.8% pass@1 and 79.2% pass@8 (compared to 29.0% and 66.4% for SFT only).

Good For

  • Applications requiring robust mathematical problem-solving and reasoning.
  • Research into reinforcement learning from human feedback (RLHF) techniques, particularly GRPO.
  • Developing agents that need to generate structured, verifiable answers for numerical tasks.

Limitations

  • While significantly improved, its absolute accuracy on complex MATH problems is still below larger models.
  • Trained and evaluated exclusively on GSM8K and MATH datasets; performance on other reasoning benchmarks is not guaranteed.

Finetunes
1 models