iiinova — where your digital self keeps evolving

I'm excited to see how the community uses these tools, and I'm looking forward to more innovations in safe and reproducible AI! Check the scorers and Weave here: 👉 wandb.me/weave_scorers 📚 A colab: wandb.me/scorers_colab

wandb.me

Local Weave Scorers | W&B Weave

Weave's local scorers are a suite of small language models that run locally on your machine with minimal latency. These models evaluate the safety and quality of your AI system’s inputs, context, and ...

100

Thomas Capelle @capetorch.bsky.social · 21/02/2025

A personal highlight was working on the Fluency Scorer powered by AnswerDotAI ModernBERT-base; we hope to move all DeBerta-powered scorers to ModernBert in the next release so we can benefit from the longer context length and training speed!

100

Thomas Capelle @capetorch.bsky.social · 21/02/2025

As part of this initiative, we also created comprehensive evaluation datasets, drawing on invaluable contributions from the open-source community. Being a reproducibility-first company, we’ve made the full recipe public, including the scorers, model weights, and the training and evaluation datasets

100

Thomas Capelle @capetorch.bsky.social · 21/02/2025

We designed these non-LLM powered scorers to leverage state-of-the-art open source models – from the PleIAI/Celadon toxicity detector to the Vectara hallucination scorer – ensuring that our AI systems are evaluated across multiple dimensions.

100

Thomas Capelle @capetorch.bsky.social · 14/02/2025

Same vibes, PR submitted, PR merged.

010

Thomas Capelle @capetorch.bsky.social · 14/02/2025

- new MacBook pro 😍 - french keyboard layout 😭

000

Reposted by Thomas Capelle

Sara Hooker @sarahooker.bsky.social · 13/02/2025

Many people have asked me about the France Action Summit. I think a summit is typically most valuable as a catalyst, not as a solution in itself. But, will share some observations.

24210

Thomas Capelle @capetorch.bsky.social · 13/02/2025

It could have been called Gulf of North America

000

Thomas Capelle @capetorch.bsky.social · 12/02/2025

This is my favorite kind of Yoga

010

Thomas Capelle @capetorch.bsky.social · 12/02/2025

I just built a CI to run an Eval of some custom LLM scorers on top of @modal-labs.bsky.social - Great to test against different GPUs - No custom runner neded on github - Fast and nice console outputs =)

020

Thomas Capelle @capetorch.bsky.social · 09/02/2025

Échalotes, tomates sèches et moules.

000

Thomas Capelle @capetorch.bsky.social · 09/02/2025

Samedi tu fais moules et frites, dimanche tu finis les moules dans un rissoto aux moules.

100

Reposted by Thomas Capelle

August J. Pollak @augustjpollak.bsky.social · 08/02/2025

Because in the United States, it’s legal to feed chicken shit to cattle. That’s why. That’s literally the reason www.telegraph.co.uk/global-healt...

Why is the USA the only country in the world with bird flu H5N1 ripping through cattle herds?

259148294217

Thomas Capelle @capetorch.bsky.social · 06/02/2025

We raised this internally! thanks for the info.

120

Thomas Capelle @capetorch.bsky.social · 05/02/2025

This budget forcing is really smart. We could do that we prefill on API models no?

100

Thomas Capelle @capetorch.bsky.social · 05/02/2025

This is getting out of hands...

050

Thomas Capelle @capetorch.bsky.social · 05/02/2025

Share your recently used Slack emojis; they tell much about your work ambiance. I like mine =)

000

Thomas Capelle @capetorch.bsky.social · 02/02/2025

Tried going skiing, traffic won 😭

000

Thomas Capelle @capetorch.bsky.social · 01/02/2025

I think the M3 pro has lower me bandwidth

010

Thomas Capelle @capetorch.bsky.social · 29/01/2025

I mostly use Claude these days, and it works very well In cursor integration. I grab o1 for more complex stuff and when I have a detailed plan and output. How is the API speed and reliability?

100

Thomas Capelle @capetorch.bsky.social · 29/01/2025

I understand the R1 hype but are you switching from Claude/o1 to it?

100

Thomas Capelle @capetorch.bsky.social · 28/01/2025

It is not there for me (on the paid cursos sub)

100

Thomas Capelle @capetorch.bsky.social · 28/01/2025

I want gemini in Cursor please.

100

Thomas Capelle @capetorch.bsky.social · 27/01/2025

Should we buy NSQ: NVDA right now? is it going to tank more?

000

Thomas Capelle @capetorch.bsky.social · 23/01/2025

Wow this is a cool resource! Way better than injecting random wikipédia titles.

010

Thomas Capelle @capetorch.bsky.social · 21/01/2025

yeah, I think I will need to do some seeding somehow

000

Thomas Capelle @capetorch.bsky.social · 21/01/2025

How do you get variance on chatGPT/Claude inference to create dataset samples? The model tends to create basically the same content over and over. I ask for a random topic and it talks almost exclusively about cats 🐱. CC @maziyarpanahi.bsky.social @teknium.bsky.social

210

Thomas Capelle @capetorch.bsky.social · 21/01/2025

The W&B Programmer agent is now topping the SWE-bench leaderboard. Shawn Lewis (co-founder of @weightsbiases.bsky.social ) has done amazing work on this. Iterating on a complex system like this wouldn't be possible without W&B Weave. Read more about the solution here: medium.com/@shawnup/the...

020