Llama 3.1 8 B Open S F T

Introduction

The Llama-3.1-8B-Open-SFT is a state-of-the-art text generation model fine-tuned from the meta-llama/Llama-3.1-8B-Instruct. It is designed to excel in tasks such as conversational interactions, question answering, and chain-of-thought reasoning, utilizing Supervised Fine-Tuning (SFT) for enhanced performance on context-sensitive and instruction-following tasks.

Architecture

The model is built on a scalable sharded architecture, distributing its 8 billion parameters over four shards. This design ensures efficient loading and deployment for large-scale applications. Key features include Chain-of-Thought (CoT) reasoning, conversational AI capabilities, and multi-purpose functionality supporting various NLP tasks such as summarization and text completion.

Training

• Base Model: meta-llama/Llama-3.1-8B
• Dataset: O1-OPEN/OpenO1-SFT, containing 77.7k samples focused on instruction-based and open-domain tasks.
• The model has been optimized for open-domain tasks through extensive supervised fine-tuning to enhance its performance across a variety of applications.

Guide: Running Locally

1. Environment Setup:

   • Install the transformers library by Hugging Face.
   • Ensure you have Python and PyTorch installed.

2. Loading the Model:

   from transformers import AutoModelForCausalLM, AutoTokenizer
   
   model_name = "prithivMLmods/Llama-3.1-8B-Open-SFT"
   tokenizer = AutoTokenizer.from_pretrained(model_name)
   model = AutoModelForCausalLM.from_pretrained(model_name)
   

3. Inference Example:

   prompt = "Explain the concept of gravity in a simple way suitable for a 10-year-old:"
   inputs = tokenizer(prompt, return_tensors="pt")
   outputs = model.generate(**inputs, max_length=150, temperature=0.7)
   
   response = tokenizer.decode(outputs[0], skip_special_tokens=True)
   print("Model Output:", response)
   

4. Hardware Recommendations:

   • Use high-performance GPUs with at least 16GB VRAM for full precision, or 8GB for quantized models.
   • Consider cloud GPU services like AWS, GCP, or Azure for scalable resources.

5. Optimization Options:

   • Utilize Safetensors for secure weight loading.
   • Apply quantization or model parallelism to optimize for resource-constrained environments.

License

The Llama-3.1-8B-Open-SFT model is released under the CreativeML Open RAIL-M license, which allows for open and flexible use while ensuring ethical and responsible deployment.