Skip to content

Quickstart

Get from zero to a reproducible fine-tuning run in under 5 minutes.

1. Install

# Core package (manifests, eval, governance, CLI)
pip install provenir

# Add the training stack
pip install "provenir[train]"

# Or install everything
pip install "provenir[all]"

Requires Python ≥ 3.11.

2. Write a config

# my_run.yaml
name: llama3-sft-v1
backend: trl
model_name_or_path: meta-llama/Llama-3.2-1B
seed: 42
max_steps: 500
batch_size: 4

peft:
  rank: 16
  alpha: 32
  target_modules: ["q_proj", "v_proj"]
  dropout: 0.05

observability_backend: wandb
observability_project: my-project

3. Prepare your dataset

Provenir uses JSONL — one record per line, each with prompt and response keys:

{"prompt": "Summarize the following article:", "response": "The article discusses..."}
{"prompt": "Translate to French:", "response": "Voici la traduction..."}

Save it as data/train.jsonl.

4. Train

provenir train my_run.yaml --dataset data/train.jsonl

This will:

  • Hash your config and dataset to produce a reproducible RunManifest
  • Run SFT training via the TRL backend with your LoRA config
  • Log metrics to W&B (or whichever backend you configured)
  • Save the manifest to artifacts/manifests/

The output looks like:

[provenir] Run ID: 3f8a1b2c-...
[provenir] Config hash: sha256:aabbcc...
[provenir] Dataset hash: sha256:112233...
[provenir] Training complete. Manifest saved to artifacts/manifests/3f8a1b2c.json

5. Evaluate

provenir eval predictions.jsonl --dataset data/eval.jsonl --metrics all

Output:

exact_match:  0.712  [0.683, 0.741]
token_f1:     0.831  [0.809, 0.853]
bleu4:        0.489  [0.461, 0.517]
rouge_l:      0.761  [0.739, 0.783]

Numbers in brackets are Wilson 95% confidence intervals.

6. Reproduce any run

provenir reproduce artifacts/manifests/3f8a1b2c.json --output reproduced/

Provenir verifies the config hash and dataset hash before starting. If anything has changed, it raises an error.

7. Run RLAIF (no human labels)

pip install "provenir[judge-anthropic,train]"

provenir rlaif my_run.yaml \
  --dataset data/train.jsonl \
  --eval-dataset data/eval.jsonl \
  --judge anthropic \
  --iterations 3

Each iteration: generate response variants → LLM judge pairwise ranking → DPO training → eval → regression gate → next iteration.

See the RLAIF guide for full configuration options.

8. Start the REST API

pip install "provenir[serve]"
provenir serve --host 0.0.0.0 --port 8000

# Submit a training job
curl -X POST http://localhost:8000/jobs/train \
  -H "Content-Type: application/json" \
  -d '{"config": {"name": "api-run", "backend": "stub"}, "records": [...]}'

# Get audit log
curl http://localhost:8000/audit

Interactive API docs at http://localhost:8000/docs.

Next Steps