distilgpt2-tiny-conversational

This model is a fine-tuned version of distilgpt2 on a parsed version of Wizard of Wikipedia. Persona alpha/beta framework designed for use with ai-msgbot. It achieves the following results on the evaluation set:

• Loss: 2.2461

Model description

• a basic dialogue model for conversation. It can be used as a chatbot.
• check out a simple demo here

Intended uses & limitations

• usage is designed for integrating with this repo: ai-msgbot
• the main specific information to know is that the model generates whole conversations between two entities, person alpha and person beta. These entity names are used functionally as custom <bos> tokens to extract when one response ends and another begins.

Training and evaluation data

• wizard of Wikipedia parsed, from parlAI

Training procedure

• deepspeed + huggingface trainer, an example notebook is in ai-msgbot

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 4
• 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: 0.05
• num_epochs: 30

Training results

Framework versions

• Transformers 4.16.1
• Pytorch 1.10.0+cu111
• Tokenizers 0.11.0