Luca Soldaini 🎀

it’s OCR week! learn how we use verifiable rewards against unit tests to improve olmOCR’s PDF understanding state of the art OCR, fully open model:

best commute on earth

which one of you im gonna have the pleasure to see at COLM???

my keystrokes go though light-up starry cable OF COURSE my code is better than yours

12+ years in this country, first time I get to wear this sticker 🗳️

wearing italian camo* at ICML *ordering an ice lattes rather than espressos at coffee shops

new @ai2.bsky.social office has something for everyone: stunning views for the outdoorsy kind, 2.5 Gbps connection at every desk for the indoor nerds

Waymo is cool but BART from SFO to downtown SF is cooler 101 can be as dark red as you want on google maps!

2025 AI hot take: everyone should use FastText more. Word embeddings are awesome.

today might be rainy, but PNW summer is already here

I've silenced all notifications on all my devices and it's truly the best thing ever ...I am considering allowing calendar notifications tho cuz I almost missed 3 meetings already 😅

two weeks traveling and I miss my mechanical keyboard so much

when someone says they wanna bring me to their favorite italian restaurant

I am still perpetually in awe that skill emergence exists in language models million of caveats but we have models that pick up capabilities from plain text??? it's so magical, I can't believe we got such treat

bluesky deserves to know we’ve adopted a cat and he’s the most handsome boy

Summary of our recommendation we submitted to White House to ensure success of open & transparent AI As a meta point, I’m very grateful to be in a position where I can put my technical expertise in the service of policy needs 🥰

"excuse me sir do you have a moment to talk about olmOCR"

if i ever find a genie lamp im gonna save a wish for “uninvent pydantic and click” 😤

in the upcoming LLMs war, i choose a neutral team

This was such a fun project to work on! We release efficient classifiers 🌐 to partition large corpora, and use them to improve sampling for LLM pretraining great work lead by @awettig.bsky.social 👇

my toxic trait is enjoying the AWS cost optimization dashboard a little too much

one benefit of open models is privacy 🔐you can run them locally & keep all your data on your local device instead of sending it to a company elsewhere congrats @soldaini.net for heavy lifting, showing our OLMoE model can run on iPhones📱

Best part of this that Luca isn’t highlighting to start is that we trained a way better OLMoE for this too. All from better annealing and post train. Didn’t need to redo pre training. Goes to show how much potential these models have! new instruct model: huggingface.co/allenai/OLMo...

🤯 Check out our new iOS OLMoE app that runs the model on-device! We also trained new OLMoE-1B-7B-0125 this time using the Tulu 3 recipe. Very exciting that RLVR improved gsm8k by almost 10 points for OLMoE 🔥 A quick 🧵

They made me do video 😬 but for a good reason! We are launching an iOS app–it runs OLMoE locally 📱 We're gonna see more on-device AI in 2025, and wanted to offer a simple way to prototype with it App: apps.apple.com/us/app/ai2-o... Code: github.com/allenai/OLMo... Blog: allenai.org/blog/olmoe-app

dreading "how was your weekend" questions don't want to tell people I spent my sunday recompiling vLLM from scratch

three things are certain in life: death, taxes, and Claude switching to concise mode during US business hours

ok but how many LLMs you know are available on BluRay

Here is Tülu 3 405B 🐫 our open-source post-training model that surpasses the performance of DeepSeek-V3! It demonstrates that our recipe, which includes RVLR scales to 405B - with performance on par with GPT-4o, & surpassing prior open-weight post-trained models of the same size including Llama 3.1.

Do people have a sandboxed code execution setup for LLM verifiers they like? My current prototype is Docker + Google Cloud Run; if I keep cache warm, startup is about 600 ms. Firewall to block any network egress from container. Would appreciate something with lower latency

calling DeepSeek R1 via APIs (tried Together AI, DeepInfra, DeepSeek APIs, our own vLLM server) is a somber reminder of how painful thinking-MoEs are at scale at best ~20TPS, while trying to generate 16k-32k per request really shows how impressive/hard O1 or O3 deployment is

it's gonna be a sad day when thinking models will figure out ASCII art, current capabilities crack me up so much:

can someone familiar with Chinese AI space map Moonshot AI/Baichuan/MiniMax/Ziphu AI vs DeepSeek/Qwen for me? What about InternLM? I know models/papers but I have no sense of relative scale of companies 🙏

just learned how gorgeous the old Seattle Times building was, so sad it got demolished for another downtown tower

Yes. arxiv.org/abs/2501.00536

kicking off 2025 with our OLMo 2 tech report while payin homage to the sequelest of sequels 🫡 🚗 2 OLMo 2 Furious 🔥 is everythin we learned since OLMo 1, with deep dives into: 🚖 stable pretrain recipe 🚔 lr anneal 🤝 data curricula 🤝 soups 🚘 tulu post-train recipe 🚜 compute infra setup 👇🧵

Everyone wants open-source language models but no one wants to lift these heavy ass weights. We just released our paper "2 OLMo 2 Furious" Can't stop us in 2025. Links below.

looking for new color themes for my terminal and it's hard to pass on this one

Anthropic/GDM could do the funniest thing and nam their next thing Chlorofluorocarbon

Anthropic, OpenAI: mid-century revival Meta, Google: post Corporate Memphis, big-tech minimalism I'm ready to give my money to the first brutalist AI endeavor 😍