DKube-instruct

DKube-instruct is a LoRA adapter on top of Qwen/Qwen2.5-0.5B-Instruct specialized for Kubernetes ops troubleshooting:

• Debug chat (diagnose + fix)
• Error → fix QA
• kubectl / YAML suggestions

Prototype. Short MPS LoRA run; not a production SRE agent.

Base model

• Base: Qwen/Qwen2.5-0.5B-Instruct (Qwen license — follow upstream terms)
• Method: LoRA (fp16) on Apple Silicon MPS — not QLoRA/bitsandbytes

Training data (locked mix)

Source	License	Notes
AnveshGummala/k8s-troubleshooting-customdsv3	Undeclared	~97 ChatML rows; experiments only — do not claim redistributable derivatives without clarifying license
jalpan04/devops-sft-dataset	Apache-2.0	Filtered to K8s / kubectl / troubleshoot / YAML-ish rows
Scoutflo/Scoutflo-SRE-Playbooks	MIT	Top ~20 K8s failure playbooks → instruction pairs
Kubernetes website debug docs + kubectl quick-ref	CC BY 4.0	Grounded QA with attribution
spacezenmasterr/k8s-sft-cmd-en	MIT	Multi-step kubectl sequences

Not used: yifeichen/k8s-troubleshooting-data, SingulioDev/varxipod-k8s-remediation.

Intended use

Ops/platform engineers wanting a tiny local K8s troubleshooting assistant. English v1.

Out of scope

No RAG product, no live cluster exec, no DPO, no MetaKube.

How to load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-0.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="float16")
model = PeftModel.from_pretrained(model, "outputs/adapter")  # or HF repo id once published

Caveats

• Synthetic / playbook-derived answers may be incomplete; verify before production changes.
• AnveshGummala license undeclared — treat that subset as non-redistributable until clarified.
• Prototype training budget (tens–hundreds of steps); quality is smoke-test level.

Citation

If you use this adapter, credit the base Qwen model and the data sources above (especially CC BY 4.0 Kubernetes docs and MIT Scoutflo playbooks).