mamba130-proteinpretrain-quinoa

Full model finetuning of Mamba-130M-HF on the "research" split (quinoa protein sequences) of GreenBeing-Proteins dataset.

Due to limits of V100 GPU, trained 510 steps x batches of 3, ~5% of the research split.

Requires GitHub main branch of Transformers (Mamba is not included in releases)

Considering training on natural language + proteins, or new "biotokens".

More details TBD

Training procedure

Notebook: https://colab.research.google.com/drive/1W1rB6rRt8krHZSVYQ_TjbnD9OwzFQeGL

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 3
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 1
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.40.0.dev0
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2