README.md

model-index:

• name: xmanii/llama-3-8b-instruct-bnb-4bit-persian description: | Model Information

  Developed by: xmanii License: Apache-2.0 Finetuned from model: unsloth/llama-3-8b-instruct-bnb-4bit

  Model Description

  This LLaMA model was fine-tuned on a unique Persian dataset of Alpaca chat conversations, consisting of approximately 8,000 rows. Our training process utilized two H100 GPUs, completing in just under 1 hour. We leveraged the power of Unsloth and Hugging Face's TRL library to accelerate our training process by 2x.

  

  Training Resources

  • 2x H100 GPUs
  • Unsloth and Hugging Face's TRL library

  Dataset

  • Unique Persian dataset of Alpaca chat conversations
  • Approximately 8,000 rows

  Open-Source Contribution

  This model is open-source, and we invite the community to use and build upon our work. The fine-tuned LLaMA model is designed to improve Persian conversation capabilities, and we hope it will contribute to the advancement of natural language processing in the Persian language.

  Using Adapters with Unsloth

  To run the model with adapters, you can use the following code:

  import torch
  from unsloth import FastLanguageModel
  from unsloth.chat_templates import get_chat_template
  
  model_save_path = "path to the download folder"  #the hugging face folder path pulled.
  
  model, tokenizer = FastLanguageModel.from_pretrained(
      model_name=model_save_path,
      max_seq_length=4096,
      load_in_4bit=True,
  )
  FastLanguageModel.for_inference(model)  # Enable native 2x faster inference
  
  tokenizer = get_chat_template(
      tokenizer,
      chat_template="llama-3",  # use the llama-3 template
      mapping={"role": "from", "content": "value", "user": "human", "assistant": "gpt"},  # mapping the messages.
  )
  
  messages = [{"from": "human", "value": "your prompt"}]#add your prompt here as human
  inputs = tokenizer.apply_chat_template(
      messages,
      tokenize=True,
      add_generation_prompt=True,  # Must add for generation
      return_tensors="pt",
  ).to("cuda")
  
  outputs = model.generate(input_ids=inputs, max_new_tokens=2048, use_cache=True)
  response = tokenizer.batch_decode(outputs, skip_special_tokens=True)
  print(response)
  

  Full 16-bit Merged Model

  For a full 16-bit merged model, please check out xmanii/Llama3-8b-simorgh-16bit.

  Future Work

  We are working on quantizing the models and bringing them to ollama.