Skip to content

quickstart.md

Latest commit

 

History

History
86 lines (58 loc) · 2.19 KB

quickstart.md

File metadata and controls

86 lines (58 loc) · 2.19 KB
Raw

Quickstart

Get from a clone to a running fine-tune in a few minutes.

1. Clone and Configure

git clone https://github.com/SahilKumar75/TuneOS
cd TuneOS
cp .env.example .env
# Add HF_TOKEN — required for gated models (Llama 3, Mistral-instruct)

2. Start Everything

With Docker (recommended)

docker-compose up -d

This brings up the Reflex UI, the FastAPI service, Redis, and a Celery worker. Open http://localhost:3000.

Without Docker

# terminal 1 — Redis
redis-server

# terminal 2 — Celery worker (this is what runs training)
celery -A workers.celery_app worker --loglevel=info

# terminal 3 — Reflex app
reflex run

The desktop launcher does this for you: when Docker is unavailable it starts Redis and the Celery worker as local subprocesses.

3. Run Your First Fine-Tune

  1. Paste a model ID — EleutherAI/pythia-410m is small (~2 GB VRAM) and good for a first run.
  2. Walk the 7-step wizard: Model → Intent → Data → Configure → Train → Results → Deploy.
  3. Upload a small instruction CSV (or generate a synthetic one in step 3), keep the defaults, and submit.

Before a job is queued the UI checks GET /api/health/workers. If no worker is alive, the job is rejected with a clear message instead of sitting in the queue indefinitely.

4. Training Modes

The wizard supports three training modes selected in step 1:

Mode Dataset shape Endpoint
SFT (default) instruction, output columns POST /api/jobs
DPO prompt, chosen, rejected columns POST /api/jobs/dpo
Knowledge Distillation instruction, output + teacher model POST /api/jobs/distill

5. Cloud GPU via Modal (Optional)

No local GPU? In Step 4 → Compute backend, pick Modal to train on a free T4. You need MODAL_TOKEN_ID and MODAL_TOKEN_SECRET in .env — see docs/deploy.md for setup.

Gated Models

Models like Llama 3 and Mistral-instruct require accepting the license on the Hugging Face Hub and a valid HF_TOKEN in .env. A missing token fails the job immediately with a clear message rather than a silent download hang.