๊ฐ€์ž… ํ›„ ์ดˆ๋Œ€ ๋งํฌ๋ฅผ ๊ณต์œ ํ•˜๋ฉด ๋™์˜์ƒ ์žฌ์ƒ ๋ฐ ์ดˆ๋Œ€ ๋ณด์ƒ์„ ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

H
@hcompany_ai
Forging Superintelligence. Putting AI to work for you, sparking the agentic AI revolution.
๊ฐ€์ž… November 2024
71 ํŒ”๋กœ์ž‰ ์ค‘    22.7K ํŒฌ
Agents are benefiting from frontier model intelligence, but are now blocked by search performance. So we developed NeoMME-Retriever, an efficient visual document retriever for Visual RAG. We contribute two models: โ€ข NeoMME, a family of 260M and 800M multimodal-native, multilingual foundation encoders trained from scratch. One bidirectional Transformer processes raw image patches and text tokens. โ€ข NeoMME-Retriever, a visual document retriever built on the NeoMME backbone. It encodes PDF page as images without OCR and returns dense and late-interaction embeddings in one pass. Both retriever model sizes sit on ViDoRe v3's model-size Pareto frontier. With matched 2048x2048 inputs on one NVIDIA L40S, the 260M model encodes about 51 pages per second, nearly 2x ColModernVBERT's 26 pages per second. Built at @hcompany_ai by @tonywu_71 and @Aurelien_L_. We released the weights under Apache 2.0 with day-0 Hugging Face Transformers support. Read @tonywu_71's thread for more details. ๐Ÿ‘‡
๋” ๋ณด๊ธฐ
๐Ÿ‘€ Meet NeoMME: a family of 260M and 800M Multimodal-Native Multilingual efficient Encoders One bidirectional Transformer processes text tokens and raw image patches, with no pretrained vision tower, text encoder, or decoder. (1/N ๐Ÿงต)
๋” ๋ณด๊ธฐ