Pyg-VN-2.7B

This model is a fine-tuned version of PygmalionAI/pygmalion-2.7b. The base model is GPT-Neo-2.7B, which was fine-tuned with a dataset of Discord/Reddit conversations by Haru in this commit and was later fine-tuned by the PygmalionAI team on various conversational data.

Training Data

The data used for training this revision is about 59M tokens (189MB) of Visual Novel data.

Intended uses and limitations

This model was trained for research and evaluation purposes to observe what effects Visual Novels would have on a chat-bot model. It has the potential to output X-rated content.

Training Procedure and Hardware

The model was fine-tuned on a single A100-80GB with the following settings:

{
    "Batch size": 1,
    "Epoch": 1,
    "FP16 Level": O2,
    "Gradient Steps": 8,
    "Learning Rate": 5e-5,
    "Adam_Beta1": 0.9,
    "Adam_Beta2": 0.98,
    "Weight Decay": 0.1,
    "Block Size": 2048
}

Training Results:

{
    "epoch": 1.0,
    "train_loss": 2.0835590305339813,
    "train_runtime": 14124.2003,
    "train_samples": 26672,
    "train_samples_per_second": 1.888,
    "train_steps_per_second": 0.236
}

Evaluation Results:

{
    "epoch": 1.0,
    "eval_accuracy": 0.556417810103348,
    "eval_loss": 2.0116419792175293,
    "eval_runtime": 317.7849,
    "eval_samples": 1648,
    "eval_samples_per_second": 5.186,
    "eval_steps_per_second": 5.186,
    "perplexity": 7.4755820257152745
}

How to use

You can use either KoboldAI or Text-Generation-WebUI to load the model.

You can also load the model directly with a pipeline for textgen:

from transformers import pipeline
generator = pipeline('text-generation', model='AlpinDale/pyg-vn-2.7b')
generator("Last summer we", do_sample=True, min_length=50)

Framework versions

• Transformers 4.28.0.dev0
• Pytorch 1.12.1+cu116
• Datasets 2.4.0
• Tokenizers 0.12.1