This model starts from Qwen/Qwen2.5-3B-Instruct, is supervised-fine-tuned on unit tests generated by Qwen2.5-Coder-7B-Instruct, and is then trained for two GRPO epochs on the Defects4J mutation subset used by RL4TG.

Revisions

main contains the final step-98 model. Every saved training checkpoint is also available as a Hugging Face revision.

Revision	Training step	Notes
`main`	98	Final model; identical weights to `checkpoint-98`
`checkpoint-10`	10	GRPO checkpoint
`checkpoint-20`	20	GRPO checkpoint
`checkpoint-30`	30	GRPO checkpoint
`checkpoint-40`	40	GRPO checkpoint
`checkpoint-50`	50	GRPO checkpoint
`checkpoint-60`	60	Highest Correct rate in the held-out checkpoint evaluation
`checkpoint-70`	70	Highest Branch Coverage in the held-out checkpoint evaluation
`checkpoint-80`	80	GRPO checkpoint
`checkpoint-90`	90	GRPO checkpoint
`checkpoint-98`	98	Final checkpoint

Loading a Checkpoint

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "tomhu/RL4TG-Qwen2.5-3B-Coder7B-Offline-SFT-GRPO"
revision = "checkpoint-60"

tokenizer = AutoTokenizer.from_pretrained(repo_id, revision=revision)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    revision=revision,
    torch_dtype="auto",
    device_map="auto",
)

Omit revision to load the final step-98 model from main.

GRPO Configuration

Training samples: 1,582 Defects4J focal methods with eligible FAIL mutants
Epochs: 2
Global prompt batch: 32
Rollouts per prompt: 8
Sampling temperature: 1.0
Learning rate: 1e-6
Data shuffle seed: 42
Maximum prompt and response lengths: 4,096 tokens each
Checkpoint interval: 10 steps

This model is part of the RL4TG experimental model collection.