Ornith-1.0-9B-KubeOps-GGUF 0.2.0

English · Português (Brasil)

Transformers model · GGUF quantizations

GGUF quantizations of leandroxk/ornith-1.0-9b-kubeops, a Brazilian Portuguese fine-tune specialized in evidence-driven Kubernetes incident diagnosis and safe, authorized recovery.

Version 0.2.0 adds a second, fully agentic operational fine-tuning stage. It emphasizes deciding whether a change is necessary, identifying the actual source of truth, executing the smallest authorized and reversible correction, observing the resulting state, and validating the affected service end to end.

See the main model card for training data, evaluation results, recommended uses, limitations, and safety guidance.

Files

File	Quantization	Approximate size	Suggested use
ornith-1.0-9b-kubeops-f16.gguf	F16	16.69 GiB	highest fidelity; requires more memory
ornith-1.0-9b-kubeops-q8.gguf	Q8_0	8.87 GiB	high fidelity with lower memory use than F16
ornith-1.0-9b-kubeops-q6_k.gguf	Q6_K	6.85 GiB	recommended balance of quality and memory use
ornith-1.0-9b-kubeops-q4_k_m.gguf	Q4_K_M	5.24 GiB	lower memory use for constrained environments

These sizes cover the model files only. Inference also requires memory for the context, KV cache, and runtime overhead.

Download

pip install -U "huggingface_hub"

hf download \
  leandroxk/ornith-1.0-9b-kubeops-GGUF \
  ornith-1.0-9b-kubeops-q6_k.gguf \
  --local-dir .

llama.cpp usage

Use a recent version of llama.cpp:

./llama-cli \
  -m ornith-1.0-9b-kubeops-q6_k.gguf \
  -cnv \
  -p "Investigate why the checkout Service has no Endpoints. Apply the minimum safe correction only if explicitly authorized."

To configure the context and GPU offload:

./llama-cli \
  -m ornith-1.0-9b-kubeops-q6_k.gguf \
  -cnv \
  -c 8192 \
  -ngl 99

The architecture declares a 262,144-token context window, but version 0.2.0 was fine-tuned with sequences of up to 8,192 tokens. Very long contexts have not been specifically evaluated for this model.

Tool use

The source model can emit XML <tool_call> blocks. The 0.2.0 training dataset used a generic run_command tool for diagnostic and corrective commands.

The host application must decide which commands are allowed, execute them, return their results to the model, and enforce authorization. Use command allowlists, execution and output limits, separate read-only and mutating permissions, redact Secrets, and require explicit approval before every mutation.

Limitations and safety

GGUF conversion does not change the limitations of the source model. The model may still invent identifiers, command output, causes, or successful recovery; select the wrong source of truth; or suggest commands that are incompatible with the cluster.

A successful command exit code is not proof of recovery. Observe the workload after a change and separately validate the Service, EndpointSlice, route, or affected application flow.

Quantization may reduce quality relative to the FP16 checkpoint. No comparative benchmark has been completed across F16, Q8_0, Q6_K, and Q4_K_M for version 0.2.0.

License

Distributed under the MIT license inherited from deepreinforce-ai/Ornith-1.0-9B.