Tom Aarsen

132 trust practitioner @tomaarsen.com · 2,631 followers
Why they matter

Practitioner with public evidence across AI research, Models & releases.

AI signals
45
past 30d
Sources
4
distinct domains
Discussions
3
past 30d
Latest signal
4d ago
View every signal from Tom Aarsen →
Sentence Transformers, SetFit & NLTK maintainer Machine Learning Engineer at 🤗 Hugging Face

Articles & links

🤗💚I'm very excited to continue our open source and open weight journey together with NVIDIA! See the announcement here: blogs.nvidia.com/blog/nvidia-...

NVIDIA to Acquire Hugging Face blogs.nvidia.com
AI Weekly's analysis →
  • NVIDIA agreed to acquire Hugging Face for $12.93 billion, with closing expected in the first half of 2027.
  • Jensen Huang pledged that NVIDIA compute will not be required to build on or deploy through the Hugging Face platform.
  • On top of the purchase price, NVIDIA will pay up to $1 billion in employee retention bonuses, per Engadget.
Read full analysis →
View on Bluesky · ♥ 27 ↻ 3 ↩ 0 · 7 from the directory shared this · 75d ago

Check out the model directly here: huggingface.co/google/embed... Or read Google's launch post: blog.google/innovation-a...

EmbeddingGemma 2: an open, lightweight multimodal embedding model blog.google
AI Weekly's analysis →
  • Google released EmbeddingGemma 2, a 740M-parameter multimodal embedding model built on Gemma 4 and licensed under Apache 2.0.
  • The model has an 8K token context window with 768-dimensional outputs truncatable to 512, 256, or 128 via Matryoshka learning.
  • Quantized on a Pixel 11 Pro, memory is roughly 191MB text-only and 567MB multimodal; MTEB Code reaches 78.68 versus 68.76 for v1.
Read full analysis →
View on Bluesky · ♥ 0 ↻ 0 ↩ 1 · 4 from the directory shared this · 42d ago

Or use the excellent Decision Index to see an extensive benchmark across 275 models: huggingface.co/spaces/multi... Some of these are now also natively supported in transformers!

Jev Decision Index - a Hugging Face Space by multimodalart huggingface.co
AI Weekly's analysis →
  • The Jev Decision Index scores decision models against a frozen suite of 132,422 requests across 37 benchmarks.
  • Nineteen benchmarks form the scored panel across five equal-weight areas, with a 0-100 mean-of-areas metric.
  • House rules ban prompt tuning, truncation, and option filtering; abstained and errored requests score zero before chance normalization.
Read full analysis →
View on Bluesky · ♥ 2 ↻ 0 ↩ 1 · 3 from the directory shared this · 7d ago

📖 The @Google EmbeddingGemma2 team has published their technical report! EmbeddingGemma2 was just 740M parameters, but its omni-modal nature and pareto-optimal performance makes it one of the most exciting releases of the last months for me. huggingface.co/papers/2610....

Paper page - EmbeddingGemma 2: Unified and Efficient Multimodal Representations huggingface.co
AI Weekly's analysis →
  • The 744M model splits parameters across a 271M text backbone, 167M vision encoder and 305M audio encoder, outputting one 768-dim space.
  • The text-only Q4_C quantized variant peaks at 191 MB on a Google Pixel 11 Pro; the full multimodal stack takes 567 MB.
  • Native audio retrieval scores 69.54 MRR@10 against 50.95 for Google's prior EG1 ASR-plus-text pipeline, with the gap widest cross-language.
Read full analysis →
View on Bluesky · ♥ 54 ↻ 6 ↩ 1 · 2 from the directory shared this · 20d ago

We at @hf.co have been working on introducing a AutoModelForDecision to the transformers library. Specifically, we want to define an *interface*, not an *implementation*, so users can integrate the input/output structure and authors keep freedom. huggingface.co/blog/decision

Decision Models in Hugging Face huggingface.co
AI Weekly's analysis →
  • Hugging Face is introducing 'decision models' that return structured decisions with probabilities instead of generating text token by token.
  • Three primitives are defined: Choice for categorical options, Score for expected value on an ordered scale, and Noul for Bernoulli yes/no answers.
  • New Transformers classes ship with a /v1/systemone serving endpoint; llama.cpp supports GGUF decision models and TRL support is listed as incoming.
Read full analysis →
View on Bluesky · ♥ 25 ↻ 6 ↩ 2 · 3 from the directory shared this · 7d ago

🤗 I've released Sentence Transformers v6.3 and wrote a blogpost to highlight its key new feature: token pooling-aware training! Train multi-vector models to preserve retrieval quality with fewer document vectors. Here's how 🧵 huggingface.co/blog/token-p...

Token Pooling-Aware Training for Multi-Vector Embedding Models huggingface.co
AI Weekly's analysis →
  • The ordinary ModernBERT multi-vector model retains 94.6% of its unpooled NanoBEIR nDCG@10 at pool_factor=4; pooling-aware training raises that to 97.8%.
  • Training applies the retrieval loss at multiple compression levels, with Ward hierarchical clustering merging similar vectors and a straight-through estimator carrying gradients back.
  • Gains widen under harsh compression: at max_tokens=4 the pooling-aware model scores 43.00 nDCG@10 versus 37.22 for the baseline, +5.78 points.
Read full analysis →
View on Bluesky · ♥ 25 ↻ 4 ↩ 3 · 2 from the directory shared this · 6d ago

🔎 Perplexity's pplx-embed-v2-context models let you retrieve small chunks with the surrounding document still in view. 0.6B & 9B, text + images, and you can index with 9B while serving queries with 0.6B. A thread on how they work 🧵 huggingface.co/collections/...

pplx-embed-v2 - a perplexity-ai Collection huggingface.co
View on Bluesky · ♥ 15 ↻ 0 ↩ 3 · 14d ago

Recent commentary

📈 New blog post: Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers. As a practical example, I finetuned a ColBERT-style model for medical retrieval. 14.5 hours on one RTX 3090, and it beats every general-purpose retriever I could find. Thread 🧵

View on Bluesky · ♥ 21 ↻ 4 ↩ 2 · 83d ago

🤗 Google Deepmind is back with EmbeddingGemma 2, which maps text (including code), images, video & audio into one shared 768-dimensional embedding space. 740M total parameters, 100+ languages, Apache 2.0 & Sentence Transformers support. Thread with highlights 🧵

View on Bluesky · ♥ 20 ↻ 5 ↩ 1 · 42d ago

This is by far the cutest robot I've ever seen. Under $400, with a camera, speaker, LiDAR, NFC, Bluetooth, Wifi, etc. + you can train it yourself with reinforcement learning. Plus it has rollerskates. It's so precious 🦆

View on Bluesky · ♥ 18 ↻ 1 ↩ 2 · 82d ago

Have you tried model merging in Sentence Transformers yet? 🤗 You can combine compatible checkpoints without another training run. Here's a 60/40 weighted average of two MiniLM models. I've had success merging checkpoints fine-tuned from the same base model.

View on Bluesky · ♥ 11 ↻ 3 ↩ 3 · 19d ago

🧵 I've just released Sentence Transformers v6.2.0! Featuring model merging with Arcee AI's mergekit, load models faster, & evaluators share work across GPUs automatically. Plus lower-memory losses, multi-vector improvements & plenty of fixes. Thread with highlights 👇

View on Bluesky · ♥ 13 ↻ 2 ↩ 2 · 27d ago

The recent Tokenizers v1 update seriously sped up tokenization for any NLP task, but now @SimonBrandeis is cooking another improvement, especially for BERT tokenizers: an additional ~5x (or ~50x total from the previous tokenizers v0.23).

View on Bluesky · ♥ 11 ↻ 3 ↩ 1 · 33d ago

Tencent just published R3-Skill, a two-stage retrieval stack purpose-built for a problem RAG-style retrievers weren't designed for: routing LLM agent skills (think Anthropic's SKILLmd format). Two 0.6B models, both Apache 2.0, one embedding model, and one reranker. 🧵

View on Bluesky · ♥ 11 ↻ 1 ↩ 1 · 126d ago

🤗 I've just released SetFit v1.2.0! SetFit trains text classifiers from a handful of labeled examples per class by fine-tuning a Sentence Transformer, no prompts or LLMs needed. v1.2 brings support for transformers v5, Sentence Transformers v6 & huggingface_hub v1. Thread 🧵

View on Bluesky · ♥ 9 ↻ 1 ↩ 1 · 74d ago

🎉 @lightonai.bsky.social just published LightOn-rerank: rerankers that score text passages or document page images against a query. Six models: Qwen3.5 at 0.8B / 2B / 4B, each in a pointwise and a generative listwise variant. Excellent for text <-> image retrieval. 🧵

View on Bluesky · ♥ 7 ↻ 1 ↩ 1 · 123d ago

💧 Liquid AI released 2 multilingual retrieval models, the first bidirectional members of the LFM family. Both 350M params, 11 languages (ar, de, en, es, fr, it, ja, ko, no, pt, sv): - LFM2.5-Embedding-350M (bi-encoder) - LFM2.5-ColBERT-350M (multi-vector, late interaction) 🧵

View on Bluesky · ♥ 5 ↻ 1 ↩ 1 · 152d ago

In Tom Aarsen's orbit

Center = Tom Aarsen. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.

Are you Tom Aarsen? Show it.

Add the Who’s Who of AI badge to your site or bio. It links back to this profile.

Listed in AI Weekly's Who's Who of AI

Markdown: [![Listed in AI Weekly's Who's Who of AI](https://aiweekly.co/modules/custom/aiweekly_whoswho/images/whoswho-badge.svg)](https://aiweekly.co/whos-who/person/tomaarsen-com)