DEPLOY.md

# DataSense HF Demo — layout

```
hf_demo/
├── app.py              # Gradio UI + examples
├── agent.py            # generate → execute loop (sandbox)
├── model_loader.py     # base + Peft adapter
├── config.py           # model IDs & env overrides
├── prompts.py          # SYSTEM_PROMPT
├── examples.py         # 6 built-in demo tasks
├── requirements.txt    # Space dependencies
├── README.md           # HF Space card (YAML frontmatter)
├── story.html          # Full project narrative + eval tables
├── eval_hackathon_benchmarks.json
├── data/
│   ├── sales.csv
│   ├── employees.csv
│   └── students.csv
└── assets/illustrations/   # story.html figures (9 PNGs)
```

## Deploy checklist

1. Create Space (Gradio, **GPU T4** minimum)
2. Upload this folder or `git push`
3. Add secret `HF_TOKEN` if the adapter repo is private
4. Set hardware to `gpu-t4` in README frontmatter
5. Smoke-test all 6 examples after build

## Model swap path

| Checkpoint | `DATASENSE_ADAPTER_MODEL` |
|------------|---------------------------|
| **SFT v1 (demo default)** | `sanjaymalladi/DataSense-Modal-E2B-SFT` |
| EVTE-STaR Micro-1 | `sanjaymalladi/DataSense-Modal-E2B-EVTE-Star-Micro1` |
| EVTE-STaR (final) | `sanjaymalladi/DataSense-Modal-E2B-EVTE-Star` |
| SFT v2 | `sanjaymalladi/DataSense-Modal-E2B-SFT-v2` |

## Project story

Open [`story.html`](story.html) for the full narrative, hackathon eval tables, and why we ship **SFT v1** on the live demo.