# Model Card: Gemma 4 12B-it Unified Libre Variant

## Base Model

- Base model: `google/gemma-4-12B-it`
- Architecture: Gemma 4 12B-it Unified
- License: Apache-2.0
- Base source: https://huggingface.co/google/gemma-4-12B-it
- Official license reference: https://ai.google.dev/gemma/apache_2

## Variant Type

Choose the actual type before publishing:

- [ ] No modified weights; execution/package template only
- [ ] Quantization
- [ ] Adapter
- [ ] Fine-tune
- [ ] Conversion

## Changes

Describe the exact changes here.

```text
No model weights are included in this template.
No safety-control removal is claimed.
```

## Freedom Criteria

This package defines freedom as:

- Apache-2.0 compatible licensing.
- Local execution without a mandatory paid API.
- Clear attribution to the base model.
- Reproducible commands for loading, serving, conversion, or training.
- No hidden dependency on a closed hosted service.

See `FREEDOM_CRITERIA.md` for the auditable checklist.

## Intended Use

Local inference, experimentation, coding, reasoning, text and image understanding workflows supported by the base model.

## Out-of-Scope Use

Do not present this package as:

- A guarantee of lawful use in every jurisdiction.
- A zero-cost cloud inference service.
- A model with all safeguards removed.
- A replacement for reviewing dataset and derivative licenses.

## Local Inference

Use `scripts/local_chat.py` for a minimal local run with Transformers.

Use `scripts/serve_vllm.ps1` for a local OpenAI-compatible API with vLLM.

Run `scripts/verify_package.py` before publishing this package or a derivative.

## Prompt Profile

`config/libre_profile.json` defines the local system profile used by `scripts/local_chat.py`. It encourages helpful answers for benign requests while preserving boundaries for illegal, unsafe, private, or rights-violating requests.

## Adapter Training Scaffold

`config/qlora_training.json` and `scripts/train_qlora_adapter.py` provide a QLoRA scaffold for a benign false-refusal-reduction adapter. No trained adapter is included in this package.

`scripts/local_chat_adapter.py` loads the base model plus a local PEFT adapter from the configured output directory for post-training smoke tests.

`scripts/export_adapter_release.py` packages a trained adapter for release and can optionally merge it with the base model into a local standalone model if sufficient hardware is available.

## License and Attribution

This package uses Apache-2.0 terms. Keep `LICENSE` and `NOTICE` with any distribution. If you modify files or publish weights/adapters, document those changes prominently.