arcee-ai/EvolKit-75K

EvolKit-75K是一个高质量的指令调优数据集，使用Arcee AI的EvolKit创建。它在训练Arcee SuperNova和INTELLECT-1等模型中发挥了关键作用。INTELLECT-1是第一个完全去中心化的大型语言模型训练项目，利用全球资源取得了显著成果。Arcee AI在微调、偏好对齐和知识蒸馏方面的贡献，使得指令调优版本的INTELLECT-1-Instruct在性能上与集中式模型如LLaMA-2竞争。在开放科学的精神下，我们发布了EvolKit-75K数据集、从Llama-3.1-405B提取的Logits、INTELLECT-1项目的数据、检查点和PRIME框架。

EvolKit-75K is a high-quality instruction tuning dataset created using Arcee AIs EvolKit. It played a key role in training models like Arcee SuperNova and INTELLECT-1. INTELLECT-1 is the first fully decentralized training project for a large language model, achieving remarkable results while utilizing resources from across the globe. With Arcee AIs contributions in fine-tuning, preference alignment, and knowledge distillation, the instruction-tuned version, INTELLECT-1-Instruct, demonstrates competitive performance with centralized models like LLaMA-2. In the spirit of open science, we’ve released the EvolKit-75K dataset, logits extracted from Llama-3.1-405B, data, checkpoints, and the PRIME framework from the INTELLECT-1 project.