Perfect! Here's a concise Model Description section for your HuggingFace model card:
Model Description
This is a fine-tuned version of Qwen2.5-3B-Instruct specialized for mathematical reasoning. The model was trained on 15,000 carefully curated samples from AceReason-1.1-SFT using a novel loss-based difficulty selection strategy.
Key Features:
- Enhanced Math Reasoning: Achieves 64.20% on MATH-500, maintaining the strong baseline while gaining new capabilities on AIME 2025 (0% → 3.33%)
- Data-Efficient Training: Selected optimal difficulty examples (25th-75th percentile by loss) from 100K samples, achieving +34.2% improvement over random selection
- Full Fine-Tuning: All 3 billion parameters updated using DeepSpeed ZeRO-3 with conservative hyperparameters (LR: 1e-5)
Architecture:
- Base: Qwen/Qwen2.5-3B-Instruct (3.09B parameters)
- Context Window: 16,384 tokens (supports long reasoning chains)
- Training Method: Supervised fine-tuning with full parameter updates
- Optimization: DeepSpeed ZeRO-3 + Flash Attention 2
Training Approach:
The model uses automated difficulty assessment based on cross-entropy loss to identify examples in the "Zone of Proximal Development" - problems where the base model demonstrates partial understanding but requires improvement. This data selection strategy proved dramatically more effective than random sampling.
Specialization Trade-off:
This model is optimized for mathematical reasoning at the cost of some general knowledge capabilities. It excels at step-by-step problem solving, mathematical proofs, and quantitative reasoning, but shows reduced performance on general science (GPQA: -10.1%) and broad knowledge tasks (MMLU: -25.9%).
Best Use Cases:
- Mathematical problem solving and tutoring
- Step-by-step reasoning tasks
- Quantitative analysis and computation
- Math competition preparation
Training Details:
- Dataset: 15K samples (loss-based selection from AceReason-100K)
- Framework: LLaMA-Factory + DeepSpeed ZeRO-3
- Epochs: 3 with cosine LR schedule and 10% warmup
