amy
@a_e_roberts
Followers
2K
Following
1K
Media
27
Statuses
198
machine learning engineer @huggingface
London, UK
Joined April 2022
After ~4 years building SOTA models & datasets, we're sharing everything we learned in ⚡The Smol Training Playbook We cover the full LLM cycle: designing ablations, choosing an architecture, curating data, post-training, and building solid infrastructure. We'll help you
36
163
1K
New contribution ! The `keypoint-matching` pipeline is available in 🤗 transformers !
5
7
37
Introducing open-Deep-Research by @huggingface ! 💥 Deep Research from @OpenAI is really good... But it's closed, as usual. > So with a team of cracked colleagues, we set ourselves a 24hours deadline to replicate and open-source Deep Research! ➡️ We built open-Deep-Research,
78
566
4K
Excellent points! Seeing this and the recent discussions about Europe’s place in AI does seem like there’s been a weird collective forgetting about @GoogleDeepMind’s importance. Especially given the ongoing importance of RL
Re: “Every major breakthrough in AI has been American”: America does itself no favors when it overestimates its specialness. Yes, the center of the AI industry is the US (California!), but many of the breakthroughs of (neural, gradient-based) AI happened elsewhere: • LSTMs,
0
0
2
QWEN2.5VL IS OUT!!! 🔥 72B is state-of-the-art against gpt-4o and Gemini in many benchmarks 🤯 legend is back > It's natively agentic for coding and computer use > Comes in 3B, 7B, 72B > Can handle 1hr+ videos > Can generate structured output > Localize and detect objects
14
111
695
The new smaller SmolVLM models just dropped, so ofc we had to train a ColPali version for them! Introducing the ColSmol family: the 500M model can retrieve documents with higher accuracy compared to the original ColPali checkpoint with about 6x less weights 🚀 (1/4 🧵)
4
38
205
Internally named "Humanity's Last Exam final final (5)"
We’re releasing Humanity’s Last Exam, a dataset with 3,000 questions developed with hundreds of subject matter experts to capture the human frontier of knowledge and reasoning. State-of-the-art AIs get <10% accuracy and are highly overconfident. @ai_risk @scaleai
0
0
7
The world doesn’t live on a pixel grid and neither should vision models! Excited to share Moving off-the-Grid (MooG): a video model w/o grid-based representations. MooG learns detached “off-the-grid tokens” that bind to (and track) scene elements as camera & content move. 🧵
10
90
759
🤓 new @NoPriorsPod on AI x the future of math with the leads of DeepMind’s AlphaProof team: @rishicomplex @LaurentSartran @tkhubert86 Topics: - why solve math with AI - what’s still hard - do we still need Terry Tao - the value of formalism Full interview 👇
1
2
20
Llama 3.2 running 100% locally in your browser on WebGPU! 🦙 Up to 85 tokens per second! ⚡️ Powered by 🤗 Transformers.js and ONNX Runtime Web. No installation required... just visit a website! Check out the demo and source code below! 👇
12
87
488
🧡to the entire 🤗 community for 1, 000, 000 models on the hub. If @OpenAI is nothing without it’s people, @huggingface is nothing without it’s community.
3
22
118
Enough people asked - we obliged- here's the entire ColPali training set: https://t.co/Av8Lm0EsY4 ! We hope this can help bootstrap some ColPali finetuning efforts and we're eager to see cool work from the community !
huggingface.co
6
49
339
Karpathy: bullish on macrodosing
Future be like tab tab tab
1
0
27
Very glad to finally be able to share what we've been working on! It's been an exciting project and I'm very proud of what our agent has reached -- a level comparable to that of a silver medallist on the IMO 2024 problems. More details in the blog post.
We’re presenting the first AI to solve International Mathematical Olympiad problems at a silver medalist level.🥈 It combines AlphaProof, a new breakthrough model for formal reasoning, and AlphaGeometry 2, an improved version of our previous system. 🧵 https://t.co/U0OFXBia8n
1
4
32
Yay! @huggingface datasets==2.20.0 added IterableDataset checkpointing support via torchdata.stateful_dataloader.StatefulDataLoader So instead of figuring out how to rewind the DL on resume, it can now be restored from a checkpoint! This is a super-useful feature: Doc:
1
10
94
Chameleon 🦎 by @Meta is now available in @huggingface transformers 😍 A multimodal model that comes in 7B and 34B sizes 🤩 But what makes this model so special? keep reading ⇣
8
51
341