pygmalion-1.3b / README.md
TearGosling's picture
Update README.md
f75fb52
metadata
license: agpl-3.0
language:
  - en
thumbnail: null
tags:
  - text generation
  - conversational
inference: false

Pygmalion 1.3B

Model description

Pymalion 1.3B is a proof-of-concept dialogue model based on EleutherAI's pythia-1.3b-deduped.

Warning: This model is NOT suitable for use by minors. It will output X-rated content under certain circumstances.

Training data

The fine-tuning dataset consisted of 56MB of dialogue data gathered from multiple sources, which includes both real and partially machine-generated conversations.

Training procedure

Fine-tuning was done using ColossalAI (specifically, with a slightly modified version of their OPT fine-tune example) for around 11.4 million tokens over 5440 steps on a single 24GB GPU. The run took just under 21 hours.

Intended use

The easy way

We provide a notebook with a Gradio UI for playing around with the model without having to manually format inputs. This notebook can be found here.

The manual way

The model can be used as a regular text generation model, but it'll perform best if the input prompt adheres to the following format:

[CHARACTER]'s Persona: [A few sentences about the character you want the model to play]

[DIALOGUE HISTORY]
You: [Your input message here]
[CHARACTER]:

Where [CHARACTER] is, as you can probably guess, the name of the character you want the model to portray, and [DIALOGUE HISTORY] is chat history so the model can have some conversational context to draw from. Ideally it'll be pairs of messages like:

[CHARACTER]: [some dialogue here]
You: [your response to the dialogue above]

Apart from chat history, you can also just add example conversations in [DIALOGUE HISTORY] to show how the character should speak - ideally at the beginning, so it doesn't get confused as to what's conversation history vs. character definition.

Known issues

  • The model can get stuck repeating certain phrases, or sometimes even entire sentences.
    • We believe this is due to that behavior being present in the training data itself, and plan to investigate and adjust accordingly for future versions.