omg

name	omg
description	Open MetaGenomic (OMG) corpus and gLM2 mixed-modality genomic language model — large pretraining data from IMG + MGnify (Tbp-scale DNA + billions of CDS). Use when: (1) Pretraining or evaluating genomic LMs, (2) Using gLM2 embeddings for function / coevolution / regulatory syntax, (3) Distinguishing ML corpora from study reanalysis. Hugging Face: tattabio/OMG, tattabio/gLM2_650M. Upstream: https://github.com/TattaBio/OMG. Route via genome-language-model.
license	MIT
category	utilities
tags	["OMG","gLM2","foundation-model","pretraining","HuggingFace"]
upstream	https://github.com/TattaBio/OMG

OMG corpus / gLM2

Dataset: https://huggingface.co/datasets/tattabio/OMG · Model: https://huggingface.co/tattabio/gLM2_650M · Code: TattaBio/OMG

Citation

Cornman, A. et al. The OMG dataset: An Open MetaGenomic corpus for mixed-modality genomic language modeling. bioRxiv https://doi.org/10.1101/2024.08.14.607850 (2024).

See also docs/references.md and metagenomics-llm.

Analytical thinking

OMG aggregates and filters metagenomic sequence from JGI IMG and EMBL MGnify into a mixed-modality corpus (amino acids for CDS, nucleic acids for intergenic regions; ~3.1 Tbp + ~3.3B CDS in the preprint) for genomic language modeling. gLM2 (Tatta Bio / Hwang et al.) uses genomic context for functional representations and can capture cross-protein coevolutionary signal better than single-protein ESM2 or short-context Evo-1 in the OMG/gLM2 analyses — still not a substitute for experimental contacts or classical taxonomy.

Use	Yes	No
Pretrain / fine-tune / embed with gLM2	✓
Replace MetaPhlAn/Kraken for your cohort		✗
Cite as “we downloaded OMG instead of ENA runs” for a clinical study		✗ without justification

Always separate ML corpus access from biological sample reanalysis (sra-ena, mgnify analyses).

How to access

# Hugging Face datasets / transformers patterns — verify current API
# huggingface-cli download tattabio/OMG
# or datasets.load_dataset("tattabio/OMG")
# gLM2_650M weights: tattabio/gLM2_650M

Pin dataset revision + model checkpoint in Methods. Expect large disk and GPU.

Decision tree

Need metagenome-scale LM resources?
├─ Pretrain corpus / gLM2 embeddings → omg
├─ DNA NT / DNABERT-2 / Caduceus / Evo 2 → genome-language-model
├─ Protein remote search only → plmsearch / deepblast
└─ Public study FASTQ → sra-ena / mgnify (not OMG)

Related skills

genome-language-model · metagenomics-llm · nucleotide-transformer · dnabert2 · caduceus · evo2 · mgnify · public-databases · plmsearch · spacedust · eggnog-mapper