metadata

license: mit
datasets:
  - chloeliu/reddit_nosleep_posts
language:
  - en
tags:
  - fun
  - horror
  - writing
widget:
  - text: '[WP] We don''t go to ravenholm anymore [RESPONSE] '
    example_title: '[WP] We don''t go to ravenholm anymore [RESPONSE] '
co2_eq_emissions:
  emissions: 60
  source: https://mlco2.github.io/impact/#compute
  training_type: fine-tuning
  geographical_location: Oregon, USA
  hardware_used: 1 T4, Google Colab

GPT-NoSleep-355m

A finetuned version of GPT2-Medium on the 'reddit-nosleep-posts' dataset. (Linked above)

TIP You can find a larger, more capable version of the model here: GPT-NoSleep-1.5b

Training Procedure

This was trained on the 'reddt-nosleep-posts' dataset, using the "HappyTransformers" library on Google Colab. This model was trained for X epochs with learning rate 1e-2.

Biases & Limitations

This likely contains the same biases and limitations as the original GPT2 that it is based on, and additionally heavy biases from the dataset. It likely will generate offensive output.

Intended Use

This model is meant for fun, nothing else.

Sample code

#Import model:
from happytransformer import HappyGeneration
happy_gen = HappyGeneration("GPT2", "DarwinAnim8or/GPT-NoSleep-355m")

#Set generation settings:
from happytransformer import GENSettings
args_top_k = GENSettingsGENSettings(no_repeat_ngram_size=3, do_sample=True, top_k=80, temperature=0.8, max_length=150, early_stopping=False)

#Generate a response:
result = happy_gen.generate_text("[WP] We don't go to the forest at night [RESPONSE] ", args=args_top_k)

print(result)
print(result.text)