🔥Telegram İfşa

when i write "we" in a paper, i'm referring to myself and the voices.

Hey MT enthusiasts, 2nd Call for Workshop proposal for EAMT 2026! Have an awesome idea related to new (and old) trends in MT you wish to discuss with amazing people? We are open to host your event! Submit by 05-11-2025: easychair.org/my/conference?… Info: eamt2026.org

Dear MT enthusiast, Here comes the 2nd Call for Tutorial proposal for the EAMT 2026! Have something amazing you wish to demonstrate to or show to the EAMT crowd? We are open to host your event! Check for more information: eamt2026.org/calls-for-tuto…

New dataset release: 🌐FineWiki This is an updated and better extracted version of Wikipedia, covering 325+ languages. Unlike the old dataset from 2023, we kept all the math content, tables, properly rendered templates, and extracted key facts. Examples and highlights below.

My tips for vLLM for offline, batched generation for things like "You are a rater. These are requirements [...]. Rate this text:": 1. tune max_num_batched_tokens/max seq; 2. use chunked prefill and prefix cache; 3. run one warmup with prefix; 4. sort by prompt length 🚀

LinkedIn with the hot (toxic) takes again. Do not forget everyone: you work to (pay to) live, not the other way around!

Why is outlines not listed as "supported" on vLLM (docs.vllm.ai/en/stable/feat…) anymore??

📢Thrilled to introduce ATLAS 🗺️: scaling laws beyond English, for pretraining, finetuning, and the curse of multilinguality. The largest public, multilingual scaling study to-date—we ran 774 exps (10M-8B params, 400+ languages) to answer: 🌍Are scaling laws different by

Q4: When should you pretrain from scratch vs finetune a multilingual checkpoint? 🌟Answer: We found compute-optimal crossover points for every model size. Rough rule of thumb: finetune if your compute budget C is < 10^10 x N ^1.54, otherwise pretrain. 8/

I haven't used Colab in a long, long time. What's the best model you can run on colab these days?

Took an image I saw on X and asked #GPT5 whether the code was correct, biasing it slightly with a distractor hint. And yes - it starts out by saying "that's incorrect" but ends with "that's correct", without admitting fault. Very interesting to see this: chatgpt.com/s/t_690366ef38…

These days, I am mostly using ~30B models in FP8 to fit well on a 48GB card. Shout out to Red Hat AI for pushing out many FP8 versions of popular models (and other quants)!

Does anyone know anyone at the Mistral AI team? For some reason they do not have the chat template in the tokenizer config, which makes it tedious to use in typical pipelines. Chat template is in their Processor class but not in Tokenizer. huggingface.co/mistralai/Mist…

!DEADLINE EXTENSION! We have decided to extend the EAMT2026 workshop proposal submission deadline with a week! New deadline: 12 November 2025! You can find the call for papers and read more on our website: eamt2026.org Submission link: easychair.org/my/conference?…