bloom-1b1-emailgen - v1

This model is a fine-tuned version of bigscience/bloom-1b1 on the postbot/multi-emails-100k dataset.

It achieves the following results on the evaluation set:

• Loss: 1.7397

Model description

More information needed

Intended uses & limitations

⚠️ this model did not have any of the original layers frozen during training ⚠️

• while this is still an area of investigation, the model likely needs to have some layers frozen during fine-tuning to retain the multilingual capabilities in balance with learning how to write emails.

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 7e-05
• train_batch_size: 2
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 64
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
1.8465	1.0	256	1.8656
1.4903	2.0	512	1.7396

details

***** eval metrics *****  

  epoch                   =        2.0  
  eval_loss               =     1.7397
  eval_runtime            = 0:04:27.41
  eval_samples            =       4216
  eval_samples_per_second =     15.766
  eval_steps_per_second   =     15.766
  perplexity              =     5.6956

Framework versions

• Transformers 4.25.0.dev0
• Pytorch 1.13.0+cu117
• Datasets 2.6.1
• Tokenizers 0.13.1