Skip to content

AGENTS.md

Latest commit

 

History

History
92 lines (66 loc) · 2.54 KB

AGENTS.md

File metadata and controls

92 lines (66 loc) · 2.54 KB

Agent guide for pi-tinker

This repo is a Pi package that helps people start fine-tuning open-source models with Tinker.

Product promise

Help a non-expert go from:

data → prepared JSONL → validation → baseline eval → smoke training → checkpoint → before/after comparison → chat in Pi

Do not turn this into a separate training framework. Keep Tinker/Tinker Cookbook as the real training layer and generate normal editable Python.

Best first commands for users

If the user has no data yet:

/tinker demo
/tinker next
/tinker doctor

If the user has CSV/JSON/JSONL/docs, prefer the managed operator first:

/tinker improve <input> --goal "what should improve" --budget demo

Then, only after the user understands API usage and evals:

/tinker improve <input> --goal "what should improve" --budget smoke --yes
/tinker improve <input> --goal "what should improve" --budget small --yes
/tinker deploy latest

Manual path:

/tinker new <input> --goal "what should improve"
/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
/tinker next

If the user only wants conversion:

/tinker prepare data.csv --out data/train.jsonl

If the user is stuck:

/tinker doctor
/skill:tinker-debug <paste error or report>

What is possible

  • Start from a zero-data demo: /tinker demo.
  • Run a managed improve loop with /tinker improve.
  • Generate app/client snippets with /tinker deploy.
  • Convert CSV, JSON, JSONL prompt/response rows, TXT/MD files, or docs directories to chat JSONL.
  • Scaffold editable Tinker Cookbook SFT scripts.
  • Validate JSONL shape and, when dependencies are installed, renderer/token masks.
  • Create and run an eval-first baseline/checkpoint comparison flow.
  • Run a 2-step smoke test before spending real compute.
  • Monitor logs and discover checkpoints.
  • Register a Tinker sampler checkpoint as a Pi model for quick chat inspection.

What is not solved here

  • It cannot create high-quality training data from nothing.
  • It is strongest for SFT; advanced RL/DPO is mostly guided by skills, not the extension wizard.
  • It does not provide production serving; checkpoint registration is for inspection inside Pi.
  • Real training requires TINKER_API_KEY and Tinker/Tinker Cookbook dependencies.

Development checks

Before committing changes:

npm test
npm pack --dry-run

The local integration test intentionally avoids real Tinker API usage.