3 ms·

Multi-token prediction models and baselines

• skilled 2y ago
  > In April we published a paper on a new training approach for better & faster LLMs using multi-token prediction. To enable further exploration by researchers, we’ve released pre-trained models for code completion using this approach. The HN discussion on the paper, Better and Faster Large Language Models via Multi-Token Prediction - https://news.ycombinator.com/item?id=40220851 https://news.ycombinator.com/item?id=40220851 - May 2024 (128 comments)