SkillGym-Qwen3.5-4B
SkillGym-Qwen3.5-4B is Qwen3.5-4B finetuned on SkillGym trajectories to use agent skills. It is one of the models released with the paper SkillGym: Training Skill-Use Agents with Automatic Verifiable Environment Generation.
An agent skill is a folder of instructions, reference documents, and scripts that an agent can consult while solving a task. This model is trained to read the relevant skill and apply it, through tool calls or shell commands, in a sandboxed workspace.
- Paper: https://arxiv.org/abs/2609.37539
- Code: https://github.com/Reason-Wang/SkillGym
- Collection: https://huggingface.co/collections/reasonwang/skillgym
Results
Scores are on a 0-100 scale, higher is better. SkillGym is task success on the SkillGym test set. SkillEval and SkillsBench report the mean and standard deviation over three runs. Skill-Use-Bench reports the skill-use (SU) score. All models are evaluated with skills available, using MiniSwe-Agent as the harness.
| Model | Size | SkillGym | SkillEval | SkillsBench | Skill-Use-Bench |
|---|---|---|---|---|---|
| MiniCPM5 | 2B | 33.2 | 63.0 ± 0.6 | 10.8 ± 2.7 | 24.1 |
| + SkillGym SFT | 2B | 33.2 | 68.1 ± 2.0 | 6.1 ± 0.8 | 44.7 |
| Ministral-3 | 8B | 17.5 | 55.7 ± 3.1 | 3.9 ± 2.7 | 4.3 |
| + SkillGym SFT | 8B | 49.5 | 77.5 ± 0.4 | 16.9 ± 1.3 | 55.1 |
| Qwen3.5 | 4B | 33.8 | 62.8 ± 1.5 | 10.1 ± 1.3 | 8.8 |
| + SkillGym SFT (this model) | 4B | 47.0 | 70.8 ± 1.6 | 14.3 ± 1.0 | 48.7 |
| Qwen3.5 | 9B | 41.3 | 65.7 ± 1.1 | 14.8 ± 3.5 | 14.8 |
| + SkillGym SFT | 9B | 59.5 | 74.7 ± 1.2 | 22.4 ± 4.2 | 49.6 |
| Qwen3.5 | 27B | 55.3 | 76.2 ± 1.1 | 32.6 ± 2.3 | 28.3 |
| + SkillGym SFT | 27B | 62.8 | 80.1 ± 1.5 | 47.4 ± 5.1 | 74.2 |
| Qwen3.5 | 122B/10B | 53.0 | 69.8 ± 1.1 | 30.1 ± 2.5 | 16.8 |
| + SkillGym SFT | 122B/10B | 65.0 | 80.0 ± 0.3 | 53.6 ± 5.2 | 72.2 |
Training
| Base model | Qwen/Qwen3.5-4B |
| Data | reasonwang/skillgym-sft, 19k verified successful trajectories from three teacher models |
| Method | Supervised finetuning, 2 epochs |
| Learning rate | 1e-5 with linear decay, AdamW |
| Batch size | 128 |
| Max sequence length | 65,536 tokens |
| Reasoning | Trained with reasoning traces |
Usage
The model keeps the chat template of its base model and can be served with any OpenAI-compatible server, for example vLLM.
vllm serve reasonwang/SkillGym-Qwen3.5-4B
In our evaluation we use temperature 0.6, top_p 0.95, top_k 20. The skill names and descriptions are listed in the system prompt, and the agent reads the skill files itself.
Limitations
- The model is an agent policy for tasks with skills and tools. It is not tuned as a general chat assistant.
- Long episodes can exhaust the context budget before the task is finished.
- Reasoning occasionally repeats itself and runs to the per-turn token limit.
Citation
@article{wang2026skillgym,
title = {{SkillGym}: Training Skill-Use Agents with Automatic Verifiable Environment Generation},
author = {Wang, Renxi and Hee, Mingshan and Koto, Fajri and Baldwin, Timothy and Li, Haonan},
journal = {arXiv preprint arXiv:2609.37539},
year = {2026}
}