SmolLM2-135M-GRPO-Student (Unquantized Reasoning Model)

This is an unquantized student reasoning model distilled from HuggingFaceTB/SmolLM2-135M-Instruct using a Two-Stage DeepSeek-R1 Architecture (Cold-Start SFT Warm-Up + Group Relative Policy Optimization).

Evaluation Metrics (Held-Out Benchmark)

• Overall Composite Score: 85.0% (+40.6% gain over baseline 44.4%)
• Code AST Syntax: 100.0% (+50.0% gain over baseline 50.0%)
• Format Delimiters: 75.0% (+62.5% gain over baseline 12.5%)
• JSON Schema Validity: 75.0% (+25.0% gain over baseline 50.0%)
• TypeSafe AI Jev Arbiter Verdict: [DEPLOY] (Fidelity Level 1)

Available Formats

• model.safetensors: Standalone unquantized Float16 PyTorch weights
• smollm2-135m-grpo-student-f16.gguf: 16-bit GGUF for llama.cpp, Ollama, and LM Studio