q0-gsm8k-1.5b-mopd
Qwen2.5-1.5B-Instruct fine-tuned with MOPD (multi-teacher on-policy distillation) on GSM8K, distilled from a q0 cyclic-trajectory GRPO mixture.
- Base:
Qwen/Qwen2.5-1.5B-Instruct(revision989aa7980e4cf806f80c7fef2b1adb7bc71aa306) - Method: MOPD, teacher_count=2 (top-2 q0 snapshots: cycle02 + cycle03, uniformly averaged), student = 4-bit QLoRA (NF4, r=8, alpha=16).
- Selection: step 192 by GSM8K validation pass@8 (256-example slice, seed 42).
- GSM8K validation: pass@1 49.07%, pass@4 76.29%, pass@8 84.77% (256-example deterministic slice).
- MATH-500 test: pass@1 16.40%, pass@4 35.00% (500 problems, four samples each).
- Format: LoRA adapter merged into 16-bit base weights (deployable directly).
Trained on an RTX 4060 Laptop (8 GB) with 4-bit QLoRA. This is the 1.5B checkpoint (the repository's 1B-scale target), not a literal 1.0B model.