This is the shortest path to see what the package does without bringing your own data.
pi install git:github.com/gvkhosla/pi-tinkerThen in Pi:
/tinker demo
/tinker next
/tinker doctor
Expected result:
- A tiny customer-support training dataset is copied into
examples/customer-support/train.jsonl. - Editable project files are generated:
train_sft.py,eval.py,eval_checkpoint.py,tinker.yaml, andnotes/plan.md. .tinker-pi/state.jsontracks your progress./tinker nextshows the next safe step./tinker doctortells you what is missing before real Tinker API usage.
With your own data, use:
/tinker new data.csv --goal "better customer support answers"
CSV shape:
question,answer
How do I cancel?,Go to Settings → Billing → Cancel subscription.
My order is late,Sorry — send us your order number and we’ll check it.Real training requires TINKER_API_KEY and tinker-cookbook installed. The recommended flow is always:
/tinker doctor
/tinker validate data/train.jsonl --model Qwen/Qwen3.5-9B-Base
/tinker eval baseline --model Qwen/Qwen3.5-9B-Base --yes
/tinker smoke train_sft.py --yes