hudsongouge

qwen35-9b-grpo-rescue

README

License: apache-2.0

Eval (test split, vLLM, max_new=12288)

Table with columns: Metric, Think-SFT, GRPO
Metric	Think-SFT	GRPO
json_ok	69.8%	81.0%
think_closed	73.0%	87.3%
mean_jaccard	0.175	0.240
hit_max	30.2%	17.5%

n=63

Usage

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "hudsongouge/qwen35-9b-grpo-rescue")

For merged NVFP4 (vLLM): hudsongouge/qwen35-9b-nvfp4-podcast-cleaner-grpo

Training notes

Method: GRPO with hard gates (think_closed ∧ json_ok) and rescue rollouts
Init: think-SFT LoRA (qwen35-9b-nvfp4-podcast-cleaner-sft-think)
100 optimizer steps; see train_meta.json

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

hudsongouge

Model Tree

Base

Qwen/Qwen3.5-9B

Adapter

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality