The source code of (Context-aware Biases for Length Extrapolation) — [ ] Cleaning codebase — [ ] Adding scripts for training ALiBi, RoPE, T5-bias — Fineweb-Edu 🔗 🤗 — Fineweb 🔗 🤗 — WikiText-103 🔗 🤗 — WikiText-2 🔗 🤗 For Hellaswag benchmark and evaluating extrapolation please use notebook. A Cable model trained on T=1024 can extrapolate on T=8192, achieving a better performance (PPL=22.22) compared to the sinusoidal model (PPL=22.81) trained on T=8192. Cable improves the model’s extrapolation ability significantly with a negligible burden in time and memory compared to the vanilla transformer. Furthermore, compared to existing RPE methods, our approach maintains nearly identical training time and GPU memory usage, while its inference overhead remains either negligible or comparable, depending on the sequence length. If you use this repository for your research or wish to refer to our positional encoding method, please use the following BibTeX entry: This repo is based on Karpathy/Build-NanoGPT. Thanks for their excellent work.
Модальности:
Генерация текста
Задача: Генерация текста
Автор: axiomlaborg
Теги: cable, feature-extraction, length-extrapolation, context-aware, positional-encoding, Cable, custom_code, en
Лайков: 3 | Загрузок: 0
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
