Kavel Rao

BS/MS from @uwcse @uwnlp

[Kavel Rao @kavel_r]

@nouhadziri Congrats Nouha!!

[Kavel Rao @kavel_r]

@liweijianglw Congrats Liwei!!!

[Kavel Rao @kavel_r]

@liweijianglw @NatMachIntell Huge congrats Liwei! Such an amazing outcome for a super cool project.

[Kavel Rao @kavel_r]

@SeungjuHan3 Congrats Seungju!

[Alex Mallen @alextmallen]

New paper!

How should we make trade-offs between the quantity and quality of labels used for eliciting knowledge from capable AI systems?

[Nouha Dziri @nouhadziri]

Happy that WildGuard got accepted at #NeurIPS2024 D&B🚀🎉🎉🎉🎉

Make your LLM safer by using our safety toolkit:

⚔️WildGuard: https://t.co/9e4HLU2W1M
🦁WildTeaming:https://t.co/1QmwdSLIld
🔧 Evaluation suite: https://t.co/1NOIBPRwwV

[Kavel Rao @kavel_r]

WildGuard is accepted to Neurips 2024 Datasets & Benchmarks!!

[Kavel Rao @kavel_r]

With WildGuard we release WildGuardMix, a carefully balanced, multi-task moderation dataset with 92K labeled examples over 13 risk categories. WildGuardMix includes 87K training examples and a high-quality moderation eval set with >5K human-annotated items for the three tasks.

[Nouha Dziri @nouhadziri]

🚀🚀Super excited that WildTeaming got accepted at #neurips2024 (main track)🎉🎉🎉🎉🎉

See you in Vancouver to discuss LLMs safety, red-teaming, reasoning and more. Also stay tuned for a safety blogpost coming out early next week🚀🚀 #LLMs

[Kavel Rao @kavel_r]

@jxmnop @srush_nlp Congrats!!

[Seungju Han @SeungjuHan3]

😂 the paper is citing our model with the wrong name (not wildcard, but wildguard🛡️) anyway, good to see more progress on moderation models!

[Kavel Rao @kavel_r]

@jxmnop Github gives a 404, is the repo private?

[Kavel Rao @kavel_r]

Finally, a huge thank you to my co-lead
@SeungjuHan3 and all our amazing collaborators @AllysonEttinger @liweijianglw @billyuchenlin @natolambert @YejinChoinka @nouhadziri !!

[Kavel Rao @kavel_r]

We show that WildGuard outperforms the strongest existing open-source moderation baselines across tasks (by up to 25.3% on refusal detection) and matches or outperforms GPT-4 by up to 4.8%.

(Adv. = adversarial; Vani. = vanilla (non-adversarial), Harm. = Harmful prompts only)

[Kavel Rao @kavel_r]

This work is part of the @allenai AI2 Safety Toolkit, check it out for more safety work:
https://t.co/a73zGqVCr8

[Nouha Dziri @nouhadziri]

🚀💫Excited to announce the **AI2 Safety Toolkit**: an open and transparent research initiative dedicated to advancing LLMs safety

https://t.co/MR3w1T4YVn

@allen_ai

[Kavel Rao @kavel_r]

With WildGuard we release WildGuardMix, a carefully balanced, multi-task moderation dataset with 92K labeled examples over 13 risk categories. WildGuardMix includes 87K training examples and a high-quality moderation eval set with >5K human-annotated items for the three tasks.

[Kavel Rao @kavel_r]

For these reasons we introduce WildGuard: a light-weight, multi-purpose moderation tool for assessing the safety of user-LLM interactions. WildGuard provides a one-stop resource for three safety moderation tasks: prompt harmfulness, response harmfulness, and response refusal.