About

Repo for supervised fine-tuning of Qwen-3.5 9b on our own DevOps dataset using QLoRa compression method and GLM's reasoning traces.

Dataset is aimed at fine-tuning a coding agent with the ability to use generic tool calls. Tested on Pi.

bf16 base weights -> 4-bit QLoRa -> LoRa adapter

LoRa is not merged into Qwen's weights, instead is distributed separately as gguf for llama.cpp to use with --lora flag

Python workflows are configured for CUDA.

Base model weights can be fetched here: https://huggingface.co/unsloth/Qwen3.5-9B

Custom evals are used to compare baseline Qwen's DevOps agentic capabilities vs the result LoRa adapter to try and ensure no degradation of important capabilities occurs.

Samplers

temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

Training requirements

About 10GB VRAM, NVidia GPU.

Measurement

On-distribution evals in evals/ (see evals/README.md): structural YAML tasks plus live-kubectl diagnostic scenarios (a kind cluster, graded by cluster end-state), driven through the model's own trained tool-loop. This is the baseline-vs-LoRA ruler. (k8s-ai-bench was retired — off-distribution: it drives kubectl-ai, not the loop the model was fine-tuned on.)