pi-tinker
Fine-tunes open-source language models using Tinker from inside the Pi agent, preparing training data, running small jobs and comparing the trained checkpoint against the original model.
pi-tinker prepares training data, writes editable Python training and evaluation scripts, runs small training jobs on Tinker, and compares a trained checkpoint with the original model, all driven from slash commands inside the Pi agent. It builds on Tinker and the Tinker Cookbook rather than implementing its own training framework.
It installs into Pi from npm or git, and a free local demo creates sample data and scripts without calling the Tinker API, so a user can see the workflow before spending an API budget. Real fine-tuning starts from CSV, JSON, JSONL, Markdown, text or a directory of documents, moves through a smoke test before a small training run, and supports a range of open models such as members of the Qwen family and OpenAI's gpt-oss, with a documented default choice for users who are unsure which model to start from.