Grpo Finetune

GRPO Fine-Tune Skill

Keys (FIREWORKS_API_KEY, FIREWORKS_ACCOUNT_ID, OPENROUTER_API_KEY) are loaded from .env in the current directory. No extra setup needed if the notebook already ran.

What you do when this skill triggers

1. Understand the task

Read the user's description. Sample 3-5 rows from their dataset (head the .jsonl) to see the prompt format and whether rows carry a gold answer field.

2. Write reward.py

Use this exact reward — schema-only, same as the notebook. Do not add value matching, ground_truth comparison, or field-level scoring. Do not modify it.

import json
from jsonschema import validate, ValidationError

SCHEMA = {
    "type": "object",
    "required": ["vendor", "date", "amount", "currency"],
    "properties": {
        "vendor":   {"type": "string"},
        "date":     {"type": "string"},
        "amount":   {"type": "number"},
        "currency": {"type": "string"},
    },
    "additionalProperties": False,
}

def score(completion: str, row=None) -> float:
    try:
        parsed = json.loads(completion.strip())
    except (json.JSONDecodeError, ValueError):
        return 0.0
    try:
        validate(instance=parsed, schema=SCHEMA)
        return 1.0
    except ValidationError:
        return 0.5

SELF_TESTS = [
    ('{"vendor": "Acme", "date": "2024-01-15", "amount": 1250.0, "currency": "USD"}', None, 1.0),
    ('{"vendor": "Acme", "date": "2024-01-15"}', None, 0.5),
    ("not json", None, 0.0),
]

The score contract is: 1.0 = valid JSON with correct schema, 0.5 = valid JSON wrong shape, 0.0 = not JSON. This is the only reward logic needed.

3. Show it and offer the edit

Show the user reward.py and say: this is what training will optimize for — edit it if your notion of "good" differs. Wait for their go-ahead.

4. Validate

$PYTHON agent-skill/grpo-finetune/generate_reward.py --validate reward.py

Must print PASS before proceeding.

5. Run the pipeline

$PYTHON agent-skill/grpo-finetune/run_pipeline.py \
    --train <path-to-train.jsonl> \
    --eval  <path-to-eval.jsonl> \
    --task  <short-task-name> \
    --output-id <model-id>

Run this in the background immediately. Relay each checkpoint to the user as it lands — print it directly in your response, do not wait to batch them:

• >>> Dataset ready · 200 prompts
• >>> Training started on Fireworks GPUs
• >>> Training complete · model deployed to ...
• >>> Fine-tuned model · X% accuracy

The pipeline automatically runs the agent demo on sample invoices at the end.

Important: training takes 30-60+ minutes. Use a timeout of at least 7200 seconds. Do not use the default 10 minute timeout.