第1部分:深入理解“多模态”
第2部分:多模态训练是什么?
Dataset | Type | Size | How | Training weight |
M3W | Interleaved image and text dataset | 43M webpages | For each webpage, they sample a random subsequence of 256 tokens and take up to the first 5 images included in the sampled sequence. | 1.0 |
ALIGN | (Image, text) pairs | 1.8B pairs | Texts are alt-texts, averaging 12 tokens/text. | 0.2 |
LTIP | (Image, text) pairs | 312M pairs | Texts are long descriptions, averaging 20.5 tokens/text. | 0.2 |
VTP | (Video, text) pairs | 27M short videos | ~22 seconds/video on average | 0.03 |
第3部分:LMM 发展趋势
| 欢迎光临 链载Ai (https://www.lianzai.com/) | Powered by Discuz! X3.5 |