Modelpublic

logan7000/cogrpo-homo-qwen25-3b-math345-groupA

sourceHugging Faceapache-2.0updated 27d agoView on Hugging Face

0likes29downloads

Model Card

cogrpo-homo-qwen25-3b-math345-groupA

Co-GRPO (homogeneous, 2× Qwen2.5-3B cross-supervised majority-vote)

field	value
Base model	`Qwen/Qwen2.5-3B`
Role	group A (Qwen2.5-3B)
Dataset	`q1716523669/MATH-Level345`
Checkpoint	best `best_model` (selected by `eval_reward`, step 110 / 136)
eval_reward (pass@1, 150-prompt val)	0.666
Training status	✅ completed
LR	3e-6, cosinewithmin_lr (min 0.1×), warmup 0.03
Effective batch	128 prompts/step × 12 generations

How to evaluate

Fully self-contained HF model (safetensors + tokenizer + config; no optimizer state). Load directly:

python

from transformers import AutoModelForCausalLM, AutoTokenizer
m = "q1716523669/cogrpo-homo-qwen25-3b-math345-groupA"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForCausalLM.from_pretrained(m, torch_dtype="bfloat16", device_map="auto")

Evaluate on MATH-500 (or the held-out 150-prompt val) with the project's standard verifier (verifiers/qwen: math_equal / grade_answer), temperature 0.6, single-sample pass@1. This matches how eval_reward above was computed.

Provenance

Part of the co-learning (cross-supervised GRPO) study. Trained on 8×H100, ZeRO-3, vLLM colocate.