Sublime
An inspiration engine for ideas
Extract reliable PDF text at 1/32 GPT-4o cost using a 7B VLM (Fully open-source)
Allen Institute for AI introduced an open-source OCR toolkit called olmOCR that extracts plain text from PDFs at over 3000 tokens/s for about 190 USD per million pages, or 1/32 GPT-4o cost—significant for large-scale document processing.___... See more
longmeidao
@lmd
if you have highly imbalanced classes, its hard to learn (very good loss from just always predicting 0).
there are many tricks from downsampling to downweighting the negative losses to things like focal loss (https://t.co/vNdGS0PcY8... See more
alex peysakhovichx.comShu
@shu
Shiyu
@shiyu
Another RL algorithm for reasoning enters the chat... Value-based Augmented Proximal Policy Optimization (VAPO)... feels like its 2018 in Deep RL all over again. Too many algorithms. https://t.co/jaQBmNF6Sv
WJ Yang
@yangwj
Q. Z
@qz