prakashacharya/nativ-macos-control-gguf

Nativ macOS Control — Q4_K_M GGUF

This repository contains the quantized deployment model for Nativ 0.3.0. It is a modified Gemma model fine-tuned to convert short macOS requests into one validated JSON tool call. It is not a general-purpose chat model and must not directly execute its output.

Artifact

• File: nativ-macos-control.Q4_K_M.gguf
• Size: 814,262,464 bytes
• SHA-256: 598159dbf7845933d365fd50a2bd0a73135b1903b01602195d8b58dfc7f2852d
• Base: google/gemma-3-1b-it
• Base revision: dcc83ea841ab6100d6b47a070329e1ba4cf78752
• Quantization: Q4_K_M via llama.cpp
• Catalog: schema 1, 56 settings and 13 native system actions
• Compatible app: Nativ 0.3.0 or later on Apple-silicon macOS 15+

The source checkpoint used for this GGUF is the nonzero LoRA checkpoint at training step 945. Its adapter SHA-256 is 341f2da4f0d09d31836073169ac99e8444cfda12306d8e00b2e3a8f7a073b170.

Training provenance

The training split contains catalog-generated request/tool-call pairs and synthetic paraphrases produced for the same fixed labels. It contains no user conversations or private user data. The frozen evaluation and hard splits were not used for training or augmentation. Training used one epoch, seed 42, LoRA rank 16/alpha 32, learning rate 2e-4, and completion-only loss on the JSON response.

The source data and scripts are in the Nativ source repository at commit 79b008b9ceb9a6c1664a91fe83f8eed3568227e4. The dataset has no separate public license; see the repository's model/data notice before reuse.

Evaluation

Evaluation used deterministic generation on the exact GGUF with llama.cpp revision a4107133a634250c8c9d888bc0bc8520dcfd6105.

Evaluation	Exact match	Parse errors	Prompts
Raw GGUF, frozen release suite	98.3%	0.0%	632
Nativ production pipeline	98.7%	0.0%	632
Raw GGUF, natural-phrasing hard set	81.2%	0.0%	32

The production pipeline deterministically routes all system_action and quit_app requests before generation, rejects generative system_action output, validates every tool name and argument against the native catalog, and requires explicit confirmation for destructive and privacy operations. Its high-risk system_action and revoke_privacy results are both 100% on the frozen suite.

Machine-readable raw and production-pipeline reports are included in this repository.

Intended use and limitations

Use this model with Nativ or an executor that applies an equivalent typed allowlist. A valid JSON call can still represent the wrong intent. The natural-phrasing hard-set result shows that the raw model should not be trusted as a standalone automation agent. Do not pass model text to a shell, do not permit unvalidated actions, and do not enable destructive operations without a separate confirmation boundary.

License and required notice

This is a modified Gemma model and is subject to the Gemma Terms of Use and Prohibited Use Policy. This repository includes a copy of the applicable agreement and prominent modified-file notice.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms