Cookbook
Cookbook
The MinT Cookbook is a separate repository that hosts longer recipe-style examples — each one is a runnable directory with a pyproject.toml, a train.py, an autoresearch.sh, and a README that explains what the experiment shows.
Browse the public cookbook on GitHub →
Available experiments
The currently maintained experiments, all running on Qwen/Qwen3-4B-Instruct-2507:
| Experiment | What it shows | Algorithm | Primary metric |
|---|---|---|---|
| chat-dpo | Pairwise chat preference DPO with held-out preference eval | DPO | eval_pair_accuracy |
| dapo-aime | Direct GRPO on DAPO-Math-17k, AIME 2024 reportable benchmark | direct GRPO | eval_accuracy |
| fingpt | FinGPT-style finance instruction tuning, Fineval anchor + sentiment SFT | LoRA SFT | eval_accuracy |
| lawbench | Full 20-task LawBench benchmark with LoRA SFT baseline | LoRA SFT | eval_lawbench_avg |
When to use the Cookbook
- You want a complete, runnable experiment rather than a snippet.
- You're looking for a baseline or a published configuration to fork.
- You need patterns beyond the four-section algorithm pages in Customize (longer-running training, evaluation harnesses, multi-stage pipelines).
How it relates to the rest of the docs
| Resource | Audience | Length |
|---|---|---|
| Get Started → Human Quickstart | First-time users | 7-step linear flow |
| Customize | Developers picking an algorithm | One page per algorithm/concept |
| mint-quickstart | First-run reproducible scripts | One script per topic |
| mint-cookbook | Researchers running full experiments | One recipe per directory |
Contributing. The cookbook accepts community contributions. Open a pull request against mint-cookbook with a new recipe directory and a README describing the experiment, the dataset, and the expected metric.