Skip to content

MaatML

MaatML takes task-specific language models from experimentation to production: prepare → train → evaluate → export → deploy, with a validator-gated data flywheel.

Install

pip install maatml
pip install "maatml[ml]"       # training stack (torch, transformers, …)

Then:

maatml --help
maatml scaffold ~/models/my-task --architecture causal_sft --name my-task
maatml validate ~/models/my-task

Lifecycle

Command Purpose
maatml prepare Build output/prepared/{train,val,test}.jsonl
maatml train Fine-tune (--smoke, --resume, --set)
maatml evaluate Score a checkpoint (--gate for thresholds)
maatml export Bundle checkpoint (+ optional GGUF/MLX) with manifest.json
maatml verify Recompute sha256 of manifest files
maatml datagen Validator-gated seed generation
maatml ingest Map / sanitize / validate external JSONL into the seed corpus
maatml runs List training runs for a model folder
maatml scaffold Create a new standalone model folder

Core owns architectures (causal_sft, seq2seq, multi_head_classifier, dpo, orpo). Model folders own validators, metrics, generators, and sanitizers via plugins.

See the Plugin author guide for registries and folder-local packages.

Docs for contributors

Local preview of this site:

pip install "maatml[docs]"
mkdocs serve

Production deploys to Cloudflare Pages via Wrangler (wrangler.toml).