Downloads · 30 days
6
2% of all-time downloads
grpo-qwen25-3b-math345
logan7000/grpo-qwen25-3b-math345
grpo-qwen25-3b-math345 is a reinforcement learning model from logan7000. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Single-model GRPO baseline trained on q1716523669/MATH-Level345 (MATH levels 3–5).
All-time downloads
257
Public
Parameters
242K
6.2 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors6.2 GB · 100%
At a glance
- Task
- Reinforcement Learning
- License
- apache-2.0
- Model type
- qwen2
- Access
- Public
- Created
- May 29, 2026
- Updated
- Aug 31, 2026
- SHA
- 55dc129f
Base models
Model card
From the Hugging Face model README
GRPO · Qwen2.5-3B · MATH-Level345
Single-model GRPO baseline trained on q1716523669/MATH-Level345 (MATH levels 3–5).
| field | value |
|---|---|
| Method | GRPO (single-model, ground-truth reward) |
| Base model | Qwen/Qwen2.5-3B |
| Dataset | q1716523669/MATH-Level345 |
| Checkpoint | best @ step 60 (this is the best_model, selected by eval_reward) |
| eval_reward (pass@1, 150-prompt val) | 0.673 |
| LR | 3e-6, cosine_with_min_lr (min 0.1×), warmup 0.03 |
| Effective batch | 128 prompts/step × 12 generations |
| Epochs | 2 (ran full 136 steps) |
Eval
This is a full HF model (safetensors + tokenizer + config), no optimizer state. Load directly:
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("q1716523669/grpo-qwen25-3b-math345")
model = AutoModelForCausalLM.from_pretrained("q1716523669/grpo-qwen25-3b-math345", torch_dtype="bfloat16", device_map="auto")
Evaluate on MATH-500 with the repo's standard verifier (verifiers/qwen, math_equal / grade_answer), temperature 0.6, single-sample pass@1.
Provenance
- run22 in the co-learning sbatch batch · 2026-05-26
- Completed 136/136 steps, exit 0.
- Task
- Reinforcement Learning
- Type
- qwen2
- License
- apache-2.0
- Languages
- en
- Created
- May 29, 2026
- Updated
- Aug 31, 2026