📢 Update: We've deepened our exploration of VLMs on Bongard Problems with more rigorous evaluations!
The best-performing model (o1) we tested solved 43 out of 100 problems - progress, but still plenty of room for improvement!
Today, we are releasing MSTS, a new Multimodal Safety Test Suite for vision-language models!
MSTS is exciting because it tests for safety risks *created by multimodality*. Each prompt consists of a text + image that *only in combination* reveal their full unsafe meaning.
🧵
📢Forget about classical #tokenizers! Embed words via sparse activations over character triplets. (1) Similar downstream performance with fewer parameters. (2) Significant improvements in cross-lingual transfer learning. Joint work with @Aleph__Alpha
👉https://t.co/iuLFQ0oGqM
I'm thrilled to announce LlavaGuard, a family of VLM-based safeguard models that offers a versatile framework for evaluating the safety compliance of visual content.
https://t.co/dlbz2XNCT9
https://t.co/2ELpgVWrM1
🚨Introducing ALERT🚨 , a large-scale safety benchmark of > 45k instructions categorized using our new taxonomy.
✨Using prior work such as Bo Decoding Trust and our Aurora-m https://t.co/fU0OoX5m2e
and @jackclarkSF's awesome Anthropic-HH.
🙏
We are also releasing the first set of strong 7B models for French, German, Spanish, and Italian and a joint EU5 one.
Both pre-trained and instruction-tuned variants!
https://t.co/gFdOqSqPkL
Today, we are announcing Occiglot!
A large-scale collaborative research collective focusing on open-source European LLMs.
We invite anybody working on multilingual datasets, benchmarks, or models to get in touch/join our discord.
https://t.co/OcT7DNM4Ky
🥳Thirteen (13!) @NeurIPSConf 2023 papers! Super proud of our Cluster of Excellence initiative RAI — Reasonable #AI 🚀 Learning+Reasoning. Composable. Human-Like. Open. Decentgralized. Together
Announcing StripedHyena 7B — an open source model using an architecture that goes beyond Transformers achieving faster performance and longer context.
It builds on the lessons learned in past year designing efficient sequence modeling architectures.
https://t.co/UGLnfz0Dma
We release LeoLM, a series of German speaking chat models based on Lama 2 7b and 13b. They preserve their ability to understand english text. We also also versions that can reply in English and in German & perform translation between these languages: https://t.co/o4n7dy0gWz
Very happy to share our latest paper on inappropriate degeneration in text-to-image models and its mitigation:
https://t.co/T2iCtHn4Yq
In total, we evaluated over 1.5M images across 11 models and different mitigation strategies.
Takeaways: 1/n 🧵
@StanfordCRFM
🙏 for pushing our work 👇 Two key lessons are: (1) having the human "artist" in the loop is advantageous, and (2) like word2vec diffusion allows for arithmetic! Here's the classic "king - male + female = queen"
HUGE THANKS to everyone who supported us! - Our LAION-5B paper has just received the Outstanding Paper Award at NeurIPS 2022. If you want to support our next large scale projects, join our Discord. 100% Open AI. From the Community, for the Community. https://t.co/soO66GEjn3