samantha-33b

Samantha

Technical notes

This model is trained on a specialized dataset and uses special sentinel tokens to demarcate conversations.

Important Note: These sentinels are similar to gpt2-style special tokens but they are <u>NOT</u> added as special tokens in the tokenizer.

Usage

For usage, you can refer to the chat.py file in this repo for an example.

Concepts

• Each conversation consists of n "sections"
• Each section can be one of:
  • me: The model
  • person: The speaker
  • situation: relevant background information to set the context of the conversation
  • thought: Thoughts generated by the model for parsing intermediate steps etc
  • information: External information added into the context by the system running the model
• The model and speaker sections can optionally include a name like me (Samantha) or person (Dmitry)

Sentinel Tokens

• <|section|> token marks the start of a "section"
• <|endsection|> token marks the end of a "section".

Example

<|section|>situation
I am talking to Diwank. I want to ask him about his food preferences.<|endsection|>
<|section|>person (Diwank)
Hey Samantha! What do you want to talk about?<|endsection|>
<|section|>me (Samantha)