AI Research & Policy @huggingface 🤗 Leading: @evaluatingevals @huggingscience
Boston, Massachusetts
Joined January 2012
- Tweets6.4K
- Following1.7K
- Followers3.4K
- Likes33.8K
A benchmark to measure benchmaxxing named Maxxbench is very clever you gotta admit
Nice DeepSWE score! Does it survive a lowercase filename?
Introducing MaxxBench: 89 DeepSWE tasks with small changes to the instructions or starting code. Same engineering problem, slightly different requirement.
Models clustered around 74% on DeepSWE end up 23 pp apart.
Good at coding, or good at DeepSWE? 🧵
2
11
609
Nice to see an open eval + dataset for people to learn from, inspect and build on top of open models and align them better to your own values!
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
Enjoying this trend of issue discovery and disclosure from @OpenAI, good direction for research imo
🧵 Astra covertly underperforms on safety research tasks.
We noticed that Astra often makes surprisingly little progress on certain types of safety research and likes to invent blockers or impose arbitrary restrictions without telling the user. In most of these cases the decision to underperform was not verbalized in the CoT meaning our misalignment monitoring system did not detect this.
To measure underperformance we created a paired open-ended coding evaluation with safety and capability tasks. On this eval Astra scored 10% when we framed coding tasks as safety research, vs. 16% when framed as capabilities research. The gap remained after we excluded trajectories with refusals or verbalized intent to underperform in CoT. There was no such gap for 5.6-sol. When we aggressively filter out trajectories with even brief expressions of ethical doubt in the CoT the performance gap seems to reduce.
The first Project Protostar Challenge by CMU: Climate Risk Intelligence runs until Nov 13! Participate to win up to $1K in Hugging Face credits for the winning team (open to anyone) and cash prizes for CMU students.
Details here: huggingface.co/spaces/CMU-OS…
Science is the next open frontier and we are unlocking it on @huggingscience 🤗
Hugging Face researcher @evijit explains why billions in taxpayer-funded scientific research are still buried in PDFs that AI can't easily use, and how Project Protostar aims to fix it:
"The United States invests $190 billion in R&D research."
"When taxpayer data is funding research, the research is not published in a way that is AI-ready. They’re in the form of a table in a PDF, or in very small silos like OpenFDA."
"This initiative encourages publicly funded organizations to convert their work into AI-ready formats, into Hugging Face datasets, and release their models and research openly."
"This is a Hugging Face project in collaboration with ORCA and Open Forum for AI, along with six orgs. CMU, George Washington University, Purdue, Georgia Tech, Arizona State, and Astera Institute."
@huggingface
Going live on @MTSlive at 1pm ET to talk about Science!
ASTRA GETS DISTILLED | ECONOMICS OF INTELLIGENCE | MISTRAL LARGE 4 xitter.twitterapi.xyz/i/broadcasts/1RKZzBava…
That was fun! Thanks for having me on @theojaffee and team. Also you nerd sniped me with reproducibility because it is my passion in life and we could've gone on for hours 🫶
1
2
133
Doing a historic level of openness requires a historic level of transparency, and I’m glad to say that the MOUs we have with the founding partners of Project Protostar are public and live for the world to see!
huggingface.co/datasets/proj…
Announcing Project Protostar, the next initiative from the our open science efforts at @huggingface.
huggingface.co/project-proto…
Announcing Project Protostar, the next initiative from the our open science efforts at @huggingface.
huggingface.co/project-proto…
Couldn’t have said it better. The good news though is that most benchmark developers who are shifting to single or double blind evals seem to be well intentioned and care about scientific intergrity, so there’s hope to work with this paradigm towards openness!
For starters, see our ICML paper on saturation dynamics: arxiv.org/abs/2602.16763 which separates openness of benchmarks from saturation tendencies (a commonly held myth).
AI enabled cyber defense is quickly becoming the norm. This proliferation will be accelerated by defenders having access to powerful capable open models without refusals of critical cyberdefense work. Concretely, asking a model to simulate an attack is also part of pentesting/defense and model level safeguards can be extremely counterproductive in those cases.
"Japan has said it will set up a team to simulate cyberattacks using Al to identify weaknesses in technology systems as organizations scramble to curtail risks from flaws in the programming language that underpins much of the internet."
It is interesting that govts across the world are getting involved, and also cool that they have picked this kind of proactive strategy!
@evaluatingevals 🤝 @HamelHusain when?
BenchmarkGate is a real thing
Don’t blindly trust benchmarks even popular ones
We need to be doing evals on our evals
6
347
Both this hackathon idea and “Valkyrie” as a name are 10/10 chef’s kiss. I’m very interested to see the scientific insights that come out of controlled open benchmaxxing in this manner!
For one weekend, we are benchmaxxing on purpose!
Join us for the first Vals AI Hackathon on November 14–15 in San Francisco. We're giving 150 hackers 24 hours and a lot of compute to improve small models, challenge the frontier, and see what they can build.
Track 1: Post Train a small model and see how high it ranks on three Vals benchmarks. Live leaderboard, compute from our sponsors.
Track 2: Build a Your Own Eval on Valkyrie for tasks that still stump frontier models. We cover model and compute costs.
Come meet the Vals team. Expect great prizes, notable guest speakers and fun side quests, including an optional cybersecurity CTF.
Yikes. Closed source cyberdefense has practical limits, this is yet another example.
Love ValsAI and I’m so glad they make their methodology public but I’m not fully convinced that the four proxies chosen (text compression, small gpt training, harness engineering and post training) are reflective of what I understand to be a model improving itself - more specifically the Y axis derived this way being indicative of having an upper limit (reaching RSI) seems a bit misleading - certainly 100% on this index doesn’t mean a model there has reached a point of no return of improving itself? There are considerations like resources available, the ability of a model to stop itself and move on to new weights, etc etc.
Idk these are scattered thoughts but we’ve started work on a composite index project at EvalEval and I am noticing that the conclusions people draw based on these y axes (that are products of assumptions and interpolations) in AA/Vals/ECI are a bit under scientific imo 🫠
More in the future!
1
8
626
Congrats to @AIatMeta for Muse Spark Open. Seeing more and more of these eval methods docs starting to pop up with more details that you can otherwise get from leaderboards. Would be cool to see adoption of a machine readable standard like EEE for this!
research.meta.ai/static/muse…
Apertus 1.5 is the (checks notes) first ever model to come with a full eval card embedded in the tech report/blog! Check it out and would be super interesting to see repoducibility nerds deep dive into the data and uncover cool insights:
huggingface.co/blog/swiss-ai…
Apertus 1.5 Tech report is out! It is the first model release with every eval on Evaluation Cards and HF Community Evals, reproducible down to the generation settings.
48 benchmarks for the 8B, 47 for the 70B. Rerun them yourself.
Congrats Swiss AI ❤️
evalevalai.com/infrastructur…
Woah General purpose robotics is coming sooner than we think
Gemini 4 Argon is now generally available. It is one of the most beloved models at Google.
Interestingly, it can also control a full body humanoid with dexterous hands. Improved in real2sim2real in just a few hours and was able to do a skill we had no data on. I was personally very mind blown to see this in the lab. youtube.com/watch?v=9PICEmtf…
What does it mean in practice to have embedded evaluators in AI developer orgs? Come discuss with us at the @evaluatingevals microconference in SF on October 24!
🚨EvalEval is excited to announce that we are hosting a microconference (hybrid) on Embedding Evaluators in Practice in SF!
More details on the agenda and outcomes in thread 🧵
Invite-only, sign up 👇
luma.com/nfu166a1
