Introducing Qwen-2.5_1.5b_MATH_GSM8K_SFT10

Evaluation (GSM8K)

Data & Training Summary

• Method: Supervised Fine-Tuning (SFT)
• Dataset: A curated subset of GSM8K with self-verified reasoning traces
• Data quality process:
  • For each GSM8K item, multiple samples were generated using base model
  • Only samples with correct final answers were retained
  • These verified CoT samples were used for SFT
• Epochs: 10
• Learning rate: 3e-6
• batch size: 4
• gradient accumulation: 4

What This Model Improves

• Stronger consistency on math reasoning chains
• More reliable extraction of correct final answers
• Stable generation of <think> and <answer> blocks

Intended Use

• math word problems
• chain-of-thought reasoning tasks

Example Inference

A conversation between User and Assistant. The User asks a question, and the Assistant solves it. The Assistant first thinks about the reasoning process in the mind and then provides the User with the answer. The reasoning process is enclosed within <think> </think> and answer is enclosed within <answer> </answer> tags, respectively, i.e., <think> reasoning process here </think> <answer> answer here </answer>.
    User: A box contains 3 red balls and 5 blue balls. If John adds 2 more red balls, how many red balls are now in the box?
    Assistant:<think>

We start with 3 red. John adds 2 more, so total = 3 + 2 = 5. </think>
<answer>5</answer>

Contact