LoRA adapter on Qwen/Qwen3-8B trained with GRPO to improve scientific-code generation — the code-writing agent in a self-refining coding/math loop. This is checkpoint 60 of the run, the checkpoint with the statistically significant gain.

On the held-out SciCode test split, sub-step pass@1 rose from 11.46% → 15.28% (+3.82 pp) over the base model (McNemar p = 0.019).


Base model	`Qwen/Qwen3-8B`
Method	GRPO with a binary correctness reward (LoRA, r = 32, α = 64)
Checkpoint	step 60 (the significant one)

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen3-8B"
repo = "lorenzocazzador/coder-grpo-qwen3-8b"

model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, repo)
tokenizer = AutoTokenizer.from_pretrained(repo)

Adapter produced for a master's thesis on self-refining coding/math agents.