Downloads · 30 days
8
16% of all-time downloads
RL4TG-Qwen3-8B-Offline-SFT-GRPO
tomhu/RL4TG-Qwen3-8B-Offline-SFT-GRPO
RL4TG-Qwen3-8B-Offline-SFT-GRPO is a text generation model from tomhu. Use it when you need the model to write or continue text. It is set up for transformers. The card lists the license as apache-2.0.
This model is a Qwen3-8B checkpoint trained first with offline supervised fine-tuning on Qwen3-14B-generated Java tests and then with mutation-aware GRPO. It is a selected release state from the GRPO run. The reposito…
All-time downloads
49
Public
Parameters
8.2B
16.4 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors16.4 GB · 100%
At a glance
- Task
- Text Generation
- Library
- transformers
- License
- apache-2.0
- Model type
- qwen3
- Access
- Public
- Created
- Jul 25, 2026
- Updated
- Jul 25, 2026
- SHA
- 267912c4
Try a prompt
Base models
Model card
From the Hugging Face model README
RL4TG Qwen3-8B Offline SFT GRPO
This model is a Qwen3-8B checkpoint trained first with offline supervised fine-tuning on Qwen3-14B-generated Java tests and then with mutation-aware GRPO. It is a selected release state from the GRPO run. The repository name intentionally does not encode an internal checkpoint number.
Training configuration
Stage 1: offline supervised fine-tuning
| Setting | Value |
|---|---|
| Student | Qwen/Qwen3-8B |
| Data generator | Qwen/Qwen3-14B |
| Objective | causal language-model supervised fine-tuning |
| Precision | BF16 |
| Learning rate | 5e-6 |
| Epochs | 1 |
| Global batch size | 32 |
| Chat template | enabled; Qwen thinking disabled |
| Seed | 42 |
Stage 2: mutation-aware GRPO
| Setting | Value |
|---|---|
| Starting policy | offline-SFT model from stage 1 |
| Training examples | 1,582 mutation-applicable Defects4J training samples |
| Algorithm | GRPO, no KL term in reward |
| Configured epochs | 2 |
| Actor optimization precision | FP32 |
| Learning rate | 1e-6 |
| Global prompt batch | 32 |
| Rollouts per prompt | 8 |
| PPO prompt groups / expanded mini batch | 4 / 32 |
| FSDP workers / per-GPU micro batch | 4 / 1 |
| Gradient accumulation | 8 |
| Temperature | 1.0 |
| Entropy coefficient | 0.01 |
| Maximum prompt/response length | 4096 / 4096 tokens |
| Data shuffle / seed | enabled / 42 |
| Chat template | enabled; Qwen thinking disabled |
The reward is -1 for an invalid generated test. A valid test receives
0.4 * line_coverage + 0.6 * branch_coverage + mutation_score. Mutation
analysis samples up to three eligible failing mutants per rollout, with a
mutation weight of 1.0. Coverage credit is gated by focal-call, assertion, and
local focal-data provenance checks.
Intended use
The model is intended for research on Java unit-test generation and reinforcement learning for software engineering. Generated tests must be compiled and executed in an isolated environment before use.
- Task
- Text Generation
- Library
- transformers
- Type
- qwen3
- License
- apache-2.0
- Created
- Jul 25, 2026
- Updated
- Jul 25, 2026