1 रिपॉजिटरी
Extending context windows beyond training length using rotary position embeddings that generalize to longer sequences.
Distinct from Context Window Extenders: Distinct from Context Window Extenders: uses rotary position embedding extrapolation for text, not overlapping video frame windows.
Explore 1 awesome GitHub repository matching graphics & multimedia · Rotary Context Extrapolators. Refine with filters or upvote what's useful.
Baichuan-7B is an open-source 7 billion parameter bilingual Transformer model designed for text generation and few-shot learning across Chinese and English. It is built on a large Transformer architecture trained on a bilingual corpus, enabling it to produce coherent text in both languages from a single model. The model incorporates several optimization techniques that distinguish it from standard large language models. It uses rotary position embeddings that can extrapolate to longer sequences than seen during training, allowing context extension beyond the original 4096-token training lengt
Extends context beyond training length by using rotary position embeddings that generalize to longer sequences.