LLAMA-3.1-ARGUNAUT-1-8B-SFT

Introduction

LLAMA-3.1-ARGUNAUT-1-8B-SFT is a fine-tuned version of the meta-llama/Llama-3.1-8B-Instruct model. It is tailored for text generation tasks and has been trained using the TRL library. The model is designed to excel in tasks involving logic, argumentation, critical thinking, and argument mapping.

Architecture

The model is based on the meta-llama/Llama-3.1-8B-Instruct architecture. It utilizes a context length of 8196 and is trained with techniques such as packing and spectrum sampling. The model has been fine-tuned using the transformers library, ensuring compatibility with related frameworks.

Training

The model was trained using Supervised Fine-Tuning (SFT) on 1 million examples for one epoch. The training procedure employed a per-device batch size of 8, gradient accumulation of 2 steps, and gradient checkpointing. The learning rate was set to 5.0e-6, following the Tülu 3 recipe, with a cosine learning rate scheduler and a warmup ratio of 0.1. Training was conducted on two H100 GPUs on the HoreKa supercomputer.

Guide: Running Locally

To run the model locally, you can use the transformers library as follows:

from transformers import pipeline

question = "Are you familiar with Argdown syntax? What's its purpose?"
generator = pipeline("text-generation", model="DebateLabKIT/Llama-3.1-Argunaut-1-8B-SFT", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

For optimal performance, consider using cloud GPUs such as NVIDIA A100 or V100 instances from providers like AWS, GCP, or Azure.

License

The model is licensed under llama3.1. For detailed license conditions, please refer to the official documentation or source repository.