zoukmar75/burotu-glimmer

Burotu methylation reader for Muse Glimmer 30B (glimmer-v5)

Trainable weights that teach Meta's open-weight Muse Glimmer 30B to read a cell-free DNA methylation profile: a methylation adapter (3,148 CpG sites -> 16 soft tokens) and QLoRA weights (rank 16), trained in two stages on simulated plasma (TCGA tumours mixed into healthy blood). The base model is not included: load it from meta-models/Muse-Glimmer-30B (revision a4e59da52a7bc87ae7251dd5545c0dd437c44b68, 4-bit nf4) and apply these weights with the code in github.com/zoukmar/burotu (burotu.muse).

Research prototype, not a medical device. Trained and tested on simulated data; the tests on real plasma are small sanity checks.

Tumour DNA in plasma (20 tissue types, held-out people and tumours)	0.3%	1%	3%	10%
Tissue of origin right first time	15%	60%	88%	92%

Files: trainable.pt (adapter + LoRA), serving.json (panel, classes, threshold, adapter settings), mu.npy (healthy background per CpG), report.json (full benchmark). Built for the Meta Global AI Developer Hackathon (2026).