# WakeelyPro — Soup config (FREE tier)
# Fine-tune Jordanian-law model on ALL laws using free Colab T4
# Docs: https://trysoup.dev
#
# Usage (free, no server needed):
#   1. npm run soup:export              -> creates data/soup/train.jsonl
#   2. Upload train.jsonl + this soup.yaml to Colab free T4
#   3. pip install "soup-cli[train]" && soup train --config soup.yaml
#   4. soup push --model ./output --repo YOUR_HF_USERNAME/wakeelypro-jordanian-law
#
# For quick local test on small data, override: --lawType rental
# All fields are the single source of truth per config/schema.py

base: Qwen/Qwen2.5-0.5B-Instruct
task: sft

data:
  train: ./data/soup/train.jsonl
  format: alpaca
  val_split: 0.1

training:
  epochs: 3
  lr: 2.0e-05
  batch_size: 1
  # Free-tier optimizations: 4-bit NF4 + layer streaming lets 7B fit 4GB,
  # but for 0.5B this just makes it even faster/cheaper on Colab free T4.
  quantization: 4bit
  stream_layers: true
  stream_source: auto
  seed: 1234
  lora:
    r: 16
    alpha: 32
    dropout: 0.05

output: ./output/wakeelypro-soup

# Optional: when you have a paid GPU, switch base to:
# base: Qwen/Qwen2.5-7B-Instruct
# and keep the rest unchanged.