Field	Value
Base model	`Qwen/Qwen2.5-3B-Instruct`
Method	OPD initialization followed by GRPO
Teacher	`Qwen/Qwen2.5-Coder-7B-Instruct during OPD`
Released checkpoint	GRPO step 98 (2 epochs)
Training data	OPD split followed by 1,582 mutation-eligible Defects4J samples
Prompt format	Qwen2.5 chat template
Global training seed	42

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tomhu/RL4TG-Qwen2.5-3B-OPD-GRPO-2-Epochs"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

The model is intended for research on Java unit-test generation. Generated tests must still be compiled and executed in the target project's own build environment.

Checkpoint revisions

The 10-step trajectory checkpoints are available as Git revisions in this repo. The main revision is the final step-98 model.

Revision	Training step	Load argument
`step-10`	10	`revision="step-10"`
`step-20`	20	`revision="step-20"`
`step-30`	30	`revision="step-30"`
`step-40`	40	`revision="step-40"`
`step-50`	50	`revision="step-50"`
`step-60`	60	`revision="step-60"`
`step-70`	70	`revision="step-70"`
`step-80`	80	`revision="step-80"`
`step-90`	90	`revision="step-90"`