README.md

---
title: intellite-500m-sft
emoji: 💬
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 5.34.2
app_file: app.py
pinned: false
---

# intellite-500M SFT — RLHF data collector

Serves the SFT-tuned **intellite 500M** model in a chat UI. Every assistant
reply gets 👍 / 👎 buttons; each rating appends one JSONL record to a local
folder that a `CommitScheduler` pushes to a dataset repo on the Hub every
5 minutes.

Weights are loaded from a bundled bf16 checkpoint (`best.pt`, ~1 GB).

Best sampling defaults are baked into the sliders:
**temp 0.7 · top-k 40 · top-p 0.7 · rep penalty 1.1** — found by grid sweep
against this checkpoint. You can override per-message via the right-side panel.

## Setup

1. **Upload the SFT checkpoint** to the Space root as `best.pt` (or set
   `INTELLITE_CKPT=/path/to/file.pt` in Settings → Variables).
2. **Create the dataset repo** `ProCreations/Intellite-storage`
   (the scheduler will auto-create it on first push too).
3. **Set `HF_TOKEN`** in Settings → Secrets — a token with **write** scope
   on the dataset repo. Without it, the Space runs but feedback only
   persists in-memory until the container restarts.
4. (Optional) Override `FEEDBACK_REPO` in Settings → Variables if you want
   to use a different dataset repo.

## Data format

Each record is a single line of JSONL in `data/data_<uuid>.jsonl` on the
dataset repo (one file per Space replica/restart):

```json
{"ts":"2026-04-25T15:23:45","system":"You are a helpful, honest, and concise assistant.","prompt_messages":[{"role":"user","content":"..."},{"role":"assistant","content":"..."},{"role":"user","content":"..."}],"response":"...","liked":true}
```

Each record is exactly `(prompt, response, reward∈{0,1})` — the shape any
preference/RL trainer expects. For DPO, group records by identical
`prompt_messages` and pair a `liked=true` response (chosen) with a
`liked=false` one (rejected). For REINFORCE/PPO, feed `liked` as a reward.

## Downloading the data

```bash
hf download ProCreations/Intellite-storage --repo-type=dataset --local-dir ./rlhf-data
```

## Hardware: ZeroGPU (half-H200, dynamic)

This Space runs on **HuggingFace ZeroGPU** — a half-H200 slice (70 GB VRAM)
is allocated on demand each time you press Send, then released when the
reply finishes. Per-message latency:

- Cold start (first message after idle): ~3–5 s of GPU queueing + ~2 s model warm
- Warm: ~5–10 s for a typical 200–400 token reply (≈80 tok/s on H200)
- Max-length 800-token reply: ~10–15 s

The `chat` function is decorated with `@spaces.GPU(duration=60)` so the
GPU stays allocated for the duration of the streamed reply, then releases.

ZeroGPU has a **per-account daily quota** (3.5 min free / 25 min PRO);
heavy users will hit a queue. Generation is otherwise free.

If the Space stalls on cold container boot, give it ~30 s — that's the
1 GB bf16 weights downloading from `ProCreations/intellite-500m-sft`.
Subsequent restarts hit the cached copy.