SmolLM3-3B-GSM8K-SFT

SmolLM3-3B-GSM8K-SFT Fine-tuned version of HuggingFaceTB/SmolLM3-3B-Base optimized for grade school math (GSM8K benchmark). Performance | Metric | Score | |--------|-------| | **GSM8K Accuracy** | **65.8%** | | Baseline (SmolLM3-3B-Base) | 23.3% | | **Improvement** | **+42.5 pp (2.8x)** | Quick Start Expected output:** Using with vLLM (Recommended for Speed) Training Details | Parameter | Value | |-----------|-------| | **Base Model** | HuggingFaceTB/SmolLM3-3B-Base | | **Training Data** | MetaMathQA (100k samples) | | **Method** | Supervised Fine-Tuning (SFT) with TRL 1.0.0 | | **Hardware** | NVIDIA H100 80GB | | **Training Time** | ~3h 16min | | **Epochs** | 1 | | **Batch Size** | 2 (effective 16 with gradient accumulation) | | **Learning Rate** | 1e-5 | | **Max Sequence Length** | 2048 | | **Optimizer** | AdamW | Chat Template This model uses the ChatML format: Training History We tried multiple approaches to improve math reasoning: | Stage | GSM8K Accuracy | Method | Notes | |-------|----------------|--------|-------| | Baseline | 23.3% | - | SmolLM3-3B-Base with no training | | SFT V1 | 59.6% | SFT 2 epochs | MetaMathQA 50k samples | | GRPO | 58% | GRPO | GSM8K train set - ineffective | | **SFT V2** | **65.8%** | SFT 1 epoch | MetaMathQA 100k samples ✓ | Key finding:** More diverse training data (100k vs 50k samples) was more effective than more epochs or GRPO reinforcement learning. Reproduction Training and evaluation scripts are available in the `training/` folder: Limitations Optimized specifically for grade school math; may not generalize to advanced mathematics Best performance with step-by-step reasoning format ending with `#### answer` Context window limited to 2048 tokens during training Citation License Apache 2.0