---
license: apache-2.0
language:
- en
base_model: unsloth/gemma-4-31B-it
tags:
- claudie
- kubernetes
- multi-cloud
- infrastructure
- text-generation
- unsloth
---

# Claudie Expert — Gemma 4 31B

Fine-tuned [Gemma 4 31B-IT](https://huggingface.co/unsloth/gemma-4-31B-it) specialized on **[Claudie](https://claudie.io/)** — an open-source platform for managing multi-cloud and hybrid-cloud Kubernetes infrastructure.

## About Claudie

[Claudie](https://claudie.io/) provisions and manages Kubernetes clusters declaratively across AWS, Azure, GCP, OCI, Hetzner, Exoscale, CloudRift, and OpenStack via the `InputManifest` CRD. Built by [Berops](https://berops.com/).

- **Project:** [github.com/berops/claudie](https://github.com/berops/claudie)
- **Documentation:** [docs.claudie.io/latest](https://docs.claudie.io/latest/)

## What this model knows

This model was fine-tuned on ~8,000 Claudie-specific Q&A conversations covering:

- Claudie's 8 microservices (Manager, Terraformer, Ansibler, Kube-Eleven, Kuber, Claudie-Operator, Autoscaler-Adapter)
- InputManifest CRD authoring for multi-cloud / GPU / autoscaling clusters
- Debugging stuck states, NATS consumer lag, Terraform state locks, WireGuard issues
- gRPC service communication, state machine, reconciliation loops
- Claudie architecture, data flow between services
- Kubernetes integration patterns

## Files in this repo

| File | Purpose |
|------|---------|
| `model-*.safetensors` | Merged bf16 weights (for vLLM, Transformers) |
| `*.Q4_K_M.gguf` | 4-bit GGUF (for Ollama, llama.cpp) |
| `chat_template.jinja` | Gemma 4 chat template |

## Usage

### vLLM
```bash
vllm serve samuelstolicny/claudie-expert-gemma4-31b --max-model-len 8192
```

### Ollama
```bash
ollama run hf.co/samuelstolicny/claudie-expert-gemma4-31b:Q4_K_M
```

### Transformers
```python
from transformers import pipeline
pipe = pipeline("text-generation", model="samuelstolicny/claudie-expert-gemma4-31b")
print(pipe([{"role": "user", "content": "How do I add a GPU node pool in Hetzner?"}]))
```

## Training

- **Base model:** `unsloth/gemma-4-31B-it`
- **Method:** LoRA bf16 (rank 64, alpha 128, all-linear)
- **Framework:** [Unsloth](https://unsloth.ai/)
- **Dataset:** 8,012 synthetic Q&A conversations generated from the Claudie codebase + docs
- **Hardware:** 1x RTX PRO 6000 96GB

## License

Apache 2.0 — same as the base model and Claudie.