I really hate some people of ML research community that really love attaching new name to existing research and calling it theirs "Let me add one line of modification and call it something brand new, so next time they will cite mine instead of the original work" + Extra

Every. Single. Time.

Thrilled our global data ecosystem audit was accepted to #ICLR2025! Empirically, we find: 1⃣ Soaring synthetic text data: ~10M tokens (pre-2018) to 100B+ (2024). 2⃣ YouTube is now 70%+ of speech/video data but could block third-party collection. 3⃣ <0.2% of data from

x.com/ficlive/status…

I just released Sentence Transformers v4.1; featuring ONNX and OpenVINO backends for rerankers offering 2-3x speedups and improved hard negatives mining which helps prepare stronger training datasets. Details in 🧵

Looks like I am not going anywhere for a long time. One of the reasons LLMs rarely ever produce usable outputs without significant guidance for my line of work.

OpenAI looking at LangChain users.

Tested this out, and there are noticeable issues above ~20 seconds of generated audio. Input for generation needs to be chunked for it to work properly. Voice cloning also likely needs fine-tuning to work out-of-distribution. Additionally, support for streaming for prod uses.

Breaking: pleias releases a new generation of small reasoning models for RAG and source synthesis. Pleias-RAG-350M and Pleias-RAG-1B come with built-in support for source citation, SOTA performance and an accuracy comparable to models ten times their size.

We shipped an alpha version of the new Surya OCR model. No hype, just facts: - 90+ languages (focus on en, romance langs, zh, ar, ja, ko) - LaTeX and formatting - Char/word/line bboxes - ~500M non-embed params - 10-20 pages/s

10B parameter DiT trained on 80M images, all owned by Freepik . Model commercially usable, raw model without distillation, open sourced. Proud to demonstrate first model-training project with our client Freepik: "F-Lite", from fal