Downloads · 30 days
5
15% of all-time downloads
hudsongouge/qwen35-4b-grpo-rescue
qwen35-4b-grpo-rescue is a reinforcement learning model from hudsongouge. Use it for the reinforcement learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.
LoRA adapter continuing from think-SFT with GRPO + rescue rollouts (DRY + rising </think bias). Does not rewrite SFT CoT.
Downloads · 30 days
5
15% of all-time downloads
All-time downloads
34
Public
Repo size
700 MB
Likes
0
Public
Click a slice to open those files.
.safetensors680 MB · 97%
From the Hugging Face model README
LoRA adapter continuing from think-SFT with GRPO + rescue rollouts (DRY + rising </think> bias).
Does not rewrite SFT CoT.
| Metric | Think-SFT | GRPO |
|---|---|---|
| json_ok | 44.4% | 84.1% |
| think_closed | 46.0% | 92.1% |
| mean_jaccard | 0.129 | 0.235 |
| hit_max | 55.6% | 14.3% |
n=63
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-4B", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "hudsongouge/qwen35-4b-grpo-rescue")
For merged NVFP4 (vLLM): hudsongouge/qwen35-4b-nvfp4-podcast-cleaner-grpo
think_closed ∧ json_ok) and rescue rolloutsqwen35-4b-nvfp4-podcast-cleaner-sft-think)train_meta.json