text-tagger-v1

Model Details

Model Description

This model identifies multiple topics related to the text in natural language. It is finetuned on youngermax/text-tagging for 3.5 epoch over ~1.3 hours on a free Kaggle P100.

• Developed by: Lincoln Maxwell
• Model type: Generative Pretrained Transformer
• Language(s) (NLP): English
• Finetuned from model: DistilGPT2

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Direct Use

<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->

input_ids = tokenizer.encode(prompt + '<|topic|>', return_tensors='pt').to('cuda')

# Generate text
output = model.generate(
  input_ids,
  max_length=1024,
  num_return_sequences=1,
  eos_token_id=tokenizer.eos_token_id,
  pad_token_id=tokenizer.eos_token_id,
  top_k=100,
  top_p=0.5,
  temperature=1
)

# Decode the output
text = tokenizer.decode(output[0], skip_special_tokens=False, early_stopping=True)
text = text[len(prompt):text.find('<|endoftext|>')]

topics = list(set(list(map(lambda x: x.strip(), text.split('<|topic|>')))[1:]))