m-a-p/OpenLLaMA-Reproduce-2041.21B

OpenLLaMA 7Bv2 Model Card

Model Description

OpenLLaMA 7Bv2 is a cutting-edge language model, trained with a focus on delivering high-quality, contextually relevant text predictions. It leverages a diverse composite dataset that includes web-crawled data, scholarly articles, and a wide range of literature and question-answer pairs to ensure broad domain coverage and applicability.

Training Data

The model was trained on a composite dataset that includes:

• Falcon refined-web dataset
• starcoder datasets
• Contributions from Wikipedia for encyclopedic knowledge
• Academic papers from arXiv for scientific understanding
• A vast collection of books spanning multiple genres
• Stack Exchange data curated by RedPajama

Training Procedure

• Learning Rate: Utilized a maximum learning rate of 3e-4 and a minimum learning rate of 3e-5.
• Batch Size: Employed a batch size of 4 million tokens, optimizing the training process for both efficiency and performance.
• Learning Rate Scheduler: The model's learning rate scheduling closely follows the strategy used in Llama2, ensuring gradual adjustments for optimal convergence.