mist-7b-sft-gutenberg-50k

This model is a fine-tuned version of
mistralai/Mistral-7B-v0.1
on a part of the train1 split of
sanps/GutenbergFictionSummaryPrepared

It achieves the following results on the evaluation set (split: "small_val"):

• Loss: 2.1665

Loss curves:

I used the https://github.com/huggingface/alignment-handbook with few minor changes.

Weirdly enough it stopped training at 0.69 epoch (no, I am not making this up!). Will have to debug why.

The training was kind of unstable in the beginning but I lowered the lr and added max_grad_norm. However I feel the lr is too low and could have got better results with higher lr.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 2
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• total_train_batch_size: 4
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
2.1619	0.69	4282	2.1664

Framework versions

• Transformers 4.35.0
• Pytorch 2.1.0+cu118
• Datasets 2.14.6
• Tokenizers 0.14.1