qwen2.5-0.5b-math-v2

A small language model fine-tuned for mathematical reasoning on GSM8K using novel techniques not found in standard approaches.

Novel Contributions

1. Reverse Curriculum Learning (RCL): Train on harder problems first
2. Self-Verification Prompting (SVP): Model checks its own work
3. Difficulty-Aware Training: Adaptive learning based on problem complexity

Performance

Metric	Score
Accuracy (Pass@1)	0.0%
Pass@k	0.0%
Majority Voting	0.0%
Consistency	0.0%
TRUE Baseline	10.0%
Improvement	-10.0%

Training Details

• Method: QLoRA + Novel Techniques
• LoRA Rank: 16
• LoRA Alpha: 32
• Preset: quick

Citation

@misc{qwen2.5_0.5b_math_v2},
  author = {Gaurav Chaudhary},
  title = {qwen2.5-0.5b-math-v2: Novel Small Model Math Reasoning},
  year = {2026},
  publisher = {HuggingFace},
  url = {https://huggingface.co/2796gauravc/qwen2.5-0.5b-math-v2}
}