phi3-kubernetes

Phi-3-mini fine-tuned on 300 Kubernetes Q&A samples from Stack Overflow using QLoRA. Built as part of an end-to-end Local Research Assistant pipeline. See the project repo for full context, including dataset preparation, evaluation results, inference server, and CI/CD setup.

Files

File	Size	Purpose
phi3_kubernetes_lora.zip	106 MB	LoRA adapter weights (rank=16, alpha=32)
phi3-kubernetes-q4_k_m.gguf	2.16 GB	Merged GGUF, 4-bit quant — recommended deployment
phi3-kubernetes-q8_0.gguf	not hosted	Benchmarked locally; regenerable with llama-quantize from the LoRA adapter

The Q8_0 quantization was benchmarked during development (see eval section below) but is not hosted here to keep the repo small. The Q4_K_M is the deployment-recommended variant: ~3.7× faster throughput at the same VRAM ceiling.

Training

Hyperparameter	Value
Base model	microsoft/Phi-3-mini-4k-instruct
Fine-tune method	QLoRA (Unsloth)
LoRA rank	16
LoRA alpha	32
Learning rate	2e-4 (cosine schedule)
Batch size	8 (effective 16 via grad accumulation)
Epochs	3
Hardware	Colab T4 (16 GB VRAM)
Dataset	mcipriano/stackoverflow-kubernetes-questions (300 samples filtered)

Evaluation

Tested on 30 held-out K8s Q&A samples against the base phi3:mini model:

Metric	Base phi3:mini	Fine-tuned phi3-kubernetes	Δ
ROUGE-L	0.1382	0.1622	+17.4%
Avg latency	18,104 ms	9,272 ms	2× faster

Quantization benchmark (RTX 3050 Laptop, 4 GB VRAM)

Quant	TTFT (ms)	Throughput (tok/s)	Peak VRAM
Q4_K_M	2,692	63.6	3.4 GB
Q8_0	2,798	17.3	3.6 GB

Q4_K_M is ~3.7× faster at the same VRAM ceiling — chosen as the deployment default.

Usage with Ollama

ollama run hf.co/shlbnrj/phi3-kubernetes:Q4_K_M

Or manually:

wget https://huggingface.co/shlbnrj/phi3-kubernetes/resolve/main/phi3-kubernetes-q4_k_m.gguf

cat > Modelfile <<'EOF'
FROM ./phi3-kubernetes-q4_k_m.gguf
PARAMETER temperature 0.4
PARAMETER num_ctx 4096
EOF

ollama create phi3-kubernetes -f Modelfile
ollama run phi3-kubernetes "What is a Pod?"

License

MIT. The base model (Phi-3) is under the Microsoft Research License.

Limitations

This is a small (3.8B parameter) model. It works well for direct K8s knowledge questions but is unreliable for multi-step tool-use scenarios (e.g., chaining a search with a Python computation). See the project's NOTES.md for detailed failure-mode analysis.