2024 · Completed
Math Reasoning Data and Qwen2-7B Fine-Tuning
Built chain-of-thought training data for math problems, fine-tuned Qwen2-7B with LoRA, and evaluated reasoning quality with an automated semantic matching pipeline.
Data preparation
- Prepared 25,000 math problem, reasoning, and answer examples and standardized the multi-turn conversation format.
- Removed duplicate samples with embedding-based similarity checks and cleaned reasoning steps for model training.
Fine-tuning and evaluation
- Fine-tuned Qwen2-7B using LoRA and developed an automated semantic matching evaluation script.
- Improved reasoning accuracy by 12% over the baseline.