What pi-tinker outputs #

This page shows what users and agents should expect before they spend real Tinker API usage.

No-data demo #

/tinker demo

Creates a small customer-support project and wizard state:

examples/customer-support/train.jsonl
data/eval.jsonl
eval.py
eval_checkpoint.py
train_sft.py
tinker.yaml
notes/plan.md
.tinker-pi/state.json

Then:

/tinker next

shows the progress checklist:

Environment ready
Training data selected
Project files created
Data validated
Baseline eval run
2-step smoke test run
Training/checkpoint available
Checkpoint eval run
Before/after compared
Checkpoint registered for chat

Managed no-API setup #

/tinker improve data.csv --goal "better support answers" --budget demo

Runs the safe setup stages only:

• converts data to chat JSONL when needed,
• writes editable training/eval files,
• runs doctor checks,
• runs lightweight dataset validation,
• stops before API usage.

Use this before smoke, small, or real budgets.

API-using budgets #

After setting TINKER_API_KEY, move up gradually:

/tinker improve data.csv --goal "better support answers" --budget smoke --yes
/tinker improve data.csv --goal "better support answers" --budget small --yes

The final report should include:

• baseline eval score,
• smoke train status,
• training metrics,
• sampler checkpoint path,
• checkpoint eval score,
• before/after delta,
• wins/regressions,
• what data to add next,
• registered checkpoint alias.

Deploy snippets #

After a checkpoint exists:

/tinker deploy latest

Writes:

deploy/<alias>/README.md
deploy/<alias>/.env.example
deploy/<alias>/python_client.py
deploy/<alias>/node_client.mjs
deploy/<alias>/fastapi_app.py

These snippets are for quick inspection/internal app testing through Tinker's OpenAI-compatible endpoint, not a full production serving stack.