# Deployment checklist

The local model directory is about 1.2 GB. Keep it out of the Space repository:
upload the weights once to a model repository, then let the Space download them
at startup.

## 1. Verify the local artifact

From the `qwen3-finetune` repository:

```bash
test -f outputs/merged-model-p6-lr5e5-r2/config.json
test -f outputs/merged-model-p6-lr5e5-r2/model.safetensors
```

Do not upload Spanish benchmark files, sealed evaluation files, training
manifests, or autoresearch results to either the model repository or the Space.

## 2. Create and upload the model repository

Authenticate with a Hugging Face token that has permission to create a model
under the Copyleft Cultivars organization. Run these commands only after
confirming the destination names:

```bash
hf auth login
hf repos create CopyleftCultivars/qwen3-v5clean --repo-type model --public
hf upload CopyleftCultivars/qwen3-v5clean \
  /home/goya/CopyLeftCultivars/qwen3-finetune/outputs/merged-model-p6-lr5e5-r2 . \
  --repo-type model
```

Record the resulting model commit SHA. That SHA becomes the Space variable
`MODEL_REVISION`.

## 3. Create the Gradio Space

```bash
hf repos create CopyleftCultivars/qwen3-v5clean-space \
  --repo-type space --space-sdk gradio --public
hf upload CopyleftCultivars/qwen3-v5clean-space \
  /home/goya/CopyLeftCultivars/qwen3-v5clean-space . \
  --repo-type space
```

Set these Space variables in the Hugging Face settings UI or through the Hub
API:

```text
MODEL_ID=CopyleftCultivars/qwen3-v5clean
MODEL_REVISION=<model commit SHA>
```

Use an `HF_TOKEN` Space secret only when the model repository is private. Never
put a token in this workspace, `README.md`, or a commit.

## 4. Smoke-test the running Space

Check startup logs for the expected model ID and revision. Ask one short
question with reasoning disabled and one with reasoning enabled. Confirm that:

- the model loads without downloading a second fallback model;
- the answer is generated from the uploaded v5clean revision;
- the UI handles an empty message and a short conversation;
- no benchmark, training, or sealed-evaluation files are mounted;
- the Space is listed under the Copyleft Cultivars organization.

This workspace only prepares the Space code. It does not authenticate to
Hugging Face, create remote repositories, upload weights, or change Space
hardware.