jayminban/RLVR-vs-SFT-Qwen2.5-1.5b-checkpoints

RLVR vs SFT on Qwen2.5-1.5b Instruct

GRPO and SFT trained checkpoint of Qwen2.5-1.5B-Instruct with GSM8K dataset.

Part of a personal project comparing RLVR vs SFT training methods.

Result: GSM8K 69.7% → 81.6% while also improving MATH (49.2% → 52.3%).

• 📊 Full results & code (GitHub)
• 🔍 Explore prompts and model responses during benchmark! (HF Spaces)