大模型
x.com ·
@aiDotEngineer
AI Engineer 大会 Vision & OCR 专场要点：文档模型、视觉 token 与数据标注
会议笔记汇总多个要点：前沿模型训练所用网页语料中印度语言占比不足 1%，Sarvam 文档模型在见像素前已读 13 万亿 token 文本、上线四个月数字化 3500 万页；一小时视频约产生百万视觉 token 而仅约 2% 有标注，Perceptron AI 用模型决定读取哪些 token；某数据实验室 PDF 决策基准上前沿模型约 30 分，Reducto 发现提供解析版页面可提升分数并减少推理 token；Merve Noyan 用大模型标注、两个小模型校验、训练检测器，成本约 3–4 美元；Meta 对视频审核压缩相似帧、缓存热门视频判定。
AI推荐理由 AI Engineer 账号整理 Vision & OCR 专场要点，包含 Sarvam 文档模型训练 13 万亿 token、Perceptron AI 的 token 选择方案、Reducto 解析提升 PDF 决策基准、Merve Noyan 低成本标注流水线、Meta 视频审核压缩策略等具体数据点。属会议笔记式转述，数据来自演讲者主张，未经独立验证。
68
浏览 390
点赞 2
评论 1
转发 0
正文
From the back row:
- Well under one percent of the web crawl frontier models learn from is in an Indian language. Sarvam's document model read thirteen trillion tokens of text before it saw a pixel, and four months after launch it is digitizing 35 million pages.
- Feed a model an hour of video and roughly a million visual tokens go in. The only ground truth is a transcript or a few labeled frames, so it learns from about two percent of them. Perceptron AI's fix is a model that decides which tokens to read.
- A frontier model scores about thirty percent on a data lab's benchmark of decisions made from PDFs. Reducto found that handing models a parsed version of the page lifted their scores and cut their reasoning tokens.
- Merve Noyan's pipeline labels images with a big open model, has two smaller models judge the boxes, and trains a detector on the survivors, for three or four dollars. Her coding agent made extra training images by flipping traffic signs left to right, until told not to.
- Meta keeps most videos out of its three agent review pipeline. It compresses similar frames, caches verdicts on viral videos, and lets creators with a strong record skip it on metadata alone.
Vision & OCR playlist:
[youtube.com/watch?v=RQi7x-…](https://www.youtube.com/watch?v=RQi7x-navxU&list=PLcfpQ4tk2k0Wp3x8nk48FTMhxa1s8AcbM)