tinyllama-mixpretrain-quinoa-sciphi

TinyLLaMA model with continued pretraining / full-model finetuning on amino acids and simulated science textbooks.

The goal is to a create models which understand amino acid sequences and natural language descriptions or Q&A.

Training data was shuffled with:

• 50% amino acid sequences / proteins from the GreenBeing research dataset (mostly quinoa)
• 50% textbook content from the SciPhi training dataset

Training procedure

CoLab notebook: https://colab.research.google.com/drive/1dah43byt-T0HQC9eCigNbxSZ8aHu6s-W?usp=sharing

To fit on an L4 GPU, it was necessary to use max_length=400 and train_batch_size=1

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• training_steps: 15000
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.19.0
• Tokenizers 0.15.2