(Not to be confused with Pygmalion 13B and Pygmalion 2 13B.)

Pygmalion 1.3B GGML

This repository contains quantized conversions of the Pygmalion 1.3B checkpoint.

For use with frontends that support GGML quantized GPT-NeoX models, such as KoboldCpp and Oobabooga (with the CTransformers loader).

Last updated on 2023-09-23.

Recommended settings:

Pygmalion 1.3B is a limited model, left in the dust by the Pygmalion project's advancements since then. Which is a shame, as it remains one of the few conversational models available for systems with less than 2GB RAM, at least before we get TinyLLaMA and quantized Phi-1.5.

Here are some tips to get the best results you can out of this model:

• Stick to a low temperature, preferably between 0.2 and 0.7.
• Keep your repetition penalty between 1.0 and 1.02. These tiny values are required for models based on Pythia Deduped.
• If using SillyTavern, follow these settings:
• You also have to keep character descriptions to a few sentences, possibly following CharacterAI's 500-character descriptions.

Notes:

• KoboldCpp [bfc696f] was tested without OpenBLAS.
• Oobabooga [895ec9d] was tested with with the --model <model> --loader ctransformers --model_type gptneox launch arguments.
• ggerganov/ggml [8ca2c19] was used for conversion and quantization.
• The original model is available at PygmalionAI/pygmalion-1.3b.
• Earlier ggmlv2 quantizations are available here.

Below is the original model card for Pygmalion 1.3B.

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.