README.md

metadata

license: apache-2.0
datasets:
  - argilla/ultrafeedback-binarized-preferences-cleaned
language:
  - en
base_model:
  - mistralai/Mistral-7B-v0.1
library_name: transformers
tags:
  - transformers

ElEmperador.

Introduction:

ElEmperador is an ORPO-based finetinue derived from the Mistral-7B-v0.1 base model.

The argilla/ultrafeedback-binarized-preferences-cleaned dataset was used, albeit a small portion was used due to GPU constraints.

Citation

[Dettmers, T., Pagnoni, A., Holtzman, A., & Zettlemoyer, L. (2023, May 23). ] https://arxiv.org/abs/2305.14314.

Evals:

BLEU:0.209

Conclusion and Model Recipe.

ORPO is a viable RLHF algorithm to improve the performance of your models than SFT finetuning. It also helps in aligning the model’s outputs more closely with human preferences, leading to more user-friendly and acceptable results.

The model recipe: [ https://github.com/ParagEkbote/El-Emperador_ModelRecipe]

Inference Script:

def generate_response(model_name, input_text, max_new_tokens=50):
    # Load the tokenizer and model from Hugging Face Hub
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForCausalLM.from_pretrained(model_name)
    
    # Tokenize the input text
    input_ids = tokenizer(input_text, return_tensors='pt').input_ids
    
    # Generate a response using the model
    with torch.no_grad():
        generated_ids = model.generate(input_ids, max_new_tokens=max_new_tokens)
    
    # Decode the generated tokens into text
    generated_text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
    
    return generated_text

if __name__ == "__main__":
    # Set the model name from Hugging Face Hub
    model_name = "AINovice2005/ElEmperador" 
    input_text = "Hello, how are you?"

    # Generate and print the model's response
    output = generate_response(model_name, input_text)
    
    print(f"Input: {input_text}")
    print(f"Output: {output}")