lora_evo_ta_all_layers_18_attention_layers

This model is a fine-tuned version of togethercomputer/evo-1-8k-base on the None dataset. It achieves the following results on the evaluation set:

• Loss: 2.8474

Model description

Trained on single ID token "5K dataset" filtered to 4k sequences (20% for test data)

lora_alpha = 64 <--------------

lora_dropout = 0.1

lora_r = 64 <---------

epochs = 3

learning rate = 3e-4

warmup_steps=500

gradient_accumulation_steps = 1

train_batch = 1

eval_batch = 1

ONLY ATTENTION LAYER <---------------------

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant
• lr_scheduler_warmup_steps: 500
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss
3.0886	0.375	1200	3.0465
3.0274	0.75	2400	2.9992
2.9835	1.125	3600	2.9622
2.9334	1.5	4800	2.9397
2.8989	1.875	6000	2.9026
2.8609	2.25	7200	2.8744
2.8413	2.625	8400	2.8584
2.8341	3.0	9600	2.8474

Framework versions

• PEFT 0.11.1
• Transformers 4.41.2
• Pytorch 2.3.0+cu121
• Datasets 2.19.2
• Tokenizers 0.19.1