qwen3-1.7b-gsm8k-sft

Qwen3-1.7B Fine-tuned for GSM8K Math Reasoning This model is a fine-tuned version of Qwen/Qwen3-1.7B optimized for mathematical reasoning on the GSM8K benchmark. Performance | Benchmark | Accuracy | Notes | |-----------|----------|-------| | **GSM8K** | **77.2%** | Grade school math (1,319 test problems) | | **MATH-500** | **55.2%** | Competition math (500 test problems) | | Baseline GSM8K | 20% | Original Qwen3-1.7B | MATH-500 Breakdown by Difficulty Level | Level | Accuracy | |-------|----------| | Level 1 (Easiest) | 86.0% | | Level 2 | 68.9% | | Level 3 | 64.8% | | Level 4 | 54.7% | | Level 5 (Hardest) | 29.1% | MATH-500 Breakdown by Subject | Subject | Accuracy | |---------|----------| | Algebra | 71.8% | | Prealgebra | 68.3% | | Number Theory | 61.3% | | Counting & Probability | 55.3% | | Geometry | 43.9% | | Precalculus | 41.1% | | Intermediate Algebra | 32.0% | Baseline Comparison | Model | GSM8K | MATH-500 | Notes | |-------|-------|----------|-------| | **This model (SFT)** | **77.2%** | 55.2% | Optimized for GSM8K | | Qwen3-1.7B (base) | ~20% | 62.0% | Pre-training only | Note: The fine-tuned model shows significant improvement on GSM8K (+57pp) but slightly lower performance on MATH-500 compared to the base model. This is expected as the training focused on GSM8K-style problems. GGUF Quantized Versions For deployment with llama.cpp, Ollama, or other GGUF-compatible runtimes: | File | Size | Description | |------|------|-------------| | qwen3-1.7b-gsm8k-q8_0.gguf | 1.8 GB | 8-bit quantized (recommended) | | qwen3-1.7b-gsm8k-f16.gguf | 3.3 GB | Full FP16 precision | Usage with Ollama Usage with llama.cpp Training Details Dataset **Size**: 247,467 examples **Sources**: GSM8K train set (7,473 examples) MetaMathQA GSM-related examples (239,994 examples) **Format**: Conversational messages with ` ... ` chain-of-thought reasoning Training Configuration **Stage 1** (2 epochs): lr=2e-5, loss 0.30 → 0.17 **Stage 2** (1 epoch): lr=5e-6, loss 0.17 → 0.167 **Batch siz