from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="Jacob-Brokloff/qwen3-0.6b-itbench-k8s-grpo", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

Training procedure

Results

Evaluated on 18 recorded investigation trajectories from seven ITBench Kubernetes incident scenarios excluded from training.

Metric	Base Qwen3-0.6B	GRPO + LoRA
Exact root-cause entity match	22.2%	38.9%
Mean entity F1	22.2%	38.9%
Valid structured output	100%	100%

The tuned policy improved from 4/18 to 7/18 exact diagnoses.

Method

For every prompt, the current policy sampled four root-cause diagnoses. A deterministic reward combined entity-set F1, an exact-match bonus, invalid entity penalties, and structured-output validation. GRPO updated a LoRA adapter on Qwen3-0.6B.

Limitations

Only seven held-out incidents were evaluated.
Multiple trajectories from one incident are not fully independent.
The model selects entities from recorded investigation evidence; it does not operate a live Kubernetes cluster.
This is a proof of concept rather than a production remediation system.

This model was trained with GRPO, a method introduced in DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.

Framework versions

TRL: 1.8.0
Transformers: 5.13.1
Pytorch: 2.13.0
Datasets: 5.0.0
Tokenizers: 0.22.2

Citations

Cite GRPO as:

bibtex

@article{shao2024deepseekmath,
    title        = {{DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models}},
    author       = {Zhihong Shao and Peiyi Wang and Qihao Zhu and Runxin Xu and Junxiao Song and Mingchuan Zhang and Y. K. Li and Y. Wu and Daya Guo},
    year         = 2024,
    eprint       = {arXiv:2402.03300},
}

Cite TRL as:

bibtex

@software{vonwerra2020trl,
  title   = {{TRL: Transformers Reinforcement Learning}},
  author  = {von Werra, Leandro and Belkada, Younes and Tunstall, Lewis and Beeching, Edward and Thrush, Tristan and Lambert, Nathan and Huang, Shengyi and Rasul, Kashif and Gallouédec, Quentin},
  license = {Apache-2.0},
  url     = {https://github.com/huggingface/trl},
  year    = {2020}
}