mGPT: fine-tune on message data - 2E

• This model is a fine-tuned version of sberbank-ai/mGPT on 80k messages. This builds on the minimum-working-example checkpoint here.
• 2E = 2 epochs

Model description

• testing if fine-tuned personality data bleeds over to other languages without being trained in them explicitly

Interesting findings thus far:

• Passing a generic word after the <name-identifier> that is in a non-English language helps ensure the model responds in the question language (see: any example).
• Model generations (in general) remain semantically consistent, even if the generations switch from <language>to English in the middle of the generated text. This demonstrates some sort of "universal concept understanding"

Usage in python

Install the transformers library if you don't have it:

pip install -U transformers

load the model into a pipeline object:

from transformers import pipeline
import torch
device = 'cuda' if torch.cuda.is_available() else 'cpu'
my_chatbot = pipeline('text-generation', 
                      'pszemraj/mGPT-Peter-2E',
                      device=0 if device == 'cuda' else -1,
                    )

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine_with_restarts
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 1 (in addition to all training on prior checkpoints)

Framework versions

• Transformers 4.18.0
• Pytorch 1.11.0+cu113
• Datasets 2.1.0
• Tokenizers 0.12.1