from transformers import AutoModel, AutoTokenizer
# You can structure your sequence with strand tokens like this or provide protein or nucleotide sequences
# as inputs.
sequence_example_1 = "<+>MLKTLMPA<->acgtacgt"
model = AutoModel.from_pretrained("nazbijari/nammu", trust_remote_code=True).cuda()
tokenizer = AutoTokenizer.from_pretrained("nazbijari/nammu", trust_remote_code=True)
encoding = tokenizer([sequence_example_1], return_tensors='pt')
out = model(encoding.input_ids.to('cuda'))
conda env create -f environment.yml
conda activate nammu
pip install -r requirements.txt
If installing mamba-ssm or causal-conv-1d breaks: Installing a direct build wheel that is specific to your hardware and software specifications with pip is what we found to work.
Nammu is a mixed-modality genomic language model trained with bidirectional Mamba-1 blocks on TattaBio's OMG database. For details please see paper here: https://www.biorxiv.org/content/10.64898/2026.07.07.736993v1
| Model | Description | Link |
|---|---|---|
nazbijari/nammu | Nammu trained with a 20K context window on proteins first for 500K steps, contigs second for 250K steps. BiMamba used is adopted from Caduceus but adds an extra residual connection outside the BiMamba block. GPU required. | https://huggingface.co/nazbijari/nammu |
Code base containing model definitions for Nammu and other models compared against
The multimodal_benchmarking_scripts folder contains the workflows used to perform benchmarking on function (deep sea metagenome KO function) and taxonomy (CAMI marine dataset).
Contains benchmarking results for all models in the paper, CAMI results, and code to run the KO benchmark task.
All training logic, can be initiated in run.sh
Executes train.py. Provides and exampe of usage.
from transformers import AutoModel, AutoTokenizer
# You can structure your sequence with strand tokens like this or provide protein or nucleotide sequences
# as inputs.
sequence_example_1 = "<+>MLKTLMPA<->acgtacgt"
model = AutoModel.from_pretrained("nazbijari/nammu", trust_remote_code=True).cuda()
tokenizer = AutoTokenizer.from_pretrained("nazbijari/nammu", trust_remote_code=True)
encoding = tokenizer([sequence_example_1], return_tensors='pt')
out = model(encoding.input_ids.to('cuda'))
conda env create -f environment.yml
conda activate nammu
pip install -r requirements.txt
If installing mamba-ssm or causal-conv-1d breaks: Installing a direct build wheel that is specific to your hardware and software specifications with pip is what we found to work.
Nammu is a mixed-modality genomic language model trained with bidirectional Mamba-1 blocks on TattaBio's OMG database. For details please see paper here: https://www.biorxiv.org/content/10.64898/2026.07.07.736993v1
| Model | Description | Link |
|---|---|---|
nazbijari/nammu | Nammu trained with a 20K context window on proteins first for 500K steps, contigs second for 250K steps. BiMamba used is adopted from Caduceus but adds an extra residual connection outside the BiMamba block. GPU required. | https://huggingface.co/nazbijari/nammu |
Code base containing model definitions for Nammu and other models compared against
The multimodal_benchmarking_scripts folder contains the workflows used to perform benchmarking on function (deep sea metagenome KO function) and taxonomy (CAMI marine dataset).
Contains benchmarking results for all models in the paper, CAMI results, and code to run the KO benchmark task.
All training logic, can be initiated in run.sh
Executes train.py. Provides and exampe of usage.