MELD

Code for MELD: Multi-Task Equilibrated Learning Detector for AI-Generated Text.

Install

pip install .

Use

With a trained local model bundle:

meld score --model model/ --file text.txt
meld score --model model/ --file text.txt --spans
meld calibrate --model model/ --input human.jsonl
meld evaluate --model model/ --input evaluation.jsonl
meld enroll --model model/ --input samples.jsonl --family new_family --output enrolled/

JSONL records contain text; calibration and evaluation also require label (0 human, 1 AI). Calibration may include register. Recalibrate after enrollment. Scores are not authorship probabilities. Span offsets refer to the returned normalized text. Weights and datasets are supplied separately.

Train

See TRAINING.md for data and settings.

torchrun --nproc_per_node=8 -m meld.train --config training.json

Run checks with pip install '.[dev]' and pytest. Each command supports --help.