TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Show HN: New and more powerful OSS hallucination detection

2 点作者 eskibars10 个月前
Hi HN!<p>At Vectara (<a href="https:&#x2F;&#x2F;vectara.com" rel="nofollow">https:&#x2F;&#x2F;vectara.com</a>) were hyper focused on providing best in class retrieval-augmented-generation (RAG). We&#x27;ve just released a new open source hallucination detection model (available on HuggingFace and Kaggle) and associated leaderboard to show which LLMs are best at producing accurate summaries. It&#x27;s far more accurate than our previous model, which has been referenced by a number of HN users here before.<p>The reason we developed&#x2F;continue developing these open source hallucination detection models is that we&#x27;ve heard from enterprises that hallucinations are one of the top items preventing them from deploying RAG applications in production. We believe by making them open source, we can further engage the community on solving this together.<p>One question that comes up frequently when we talk about this model is &quot;how can we detect the &#x27;truthiness&#x27; of a LLM output?&quot; The answer is that our model is hyper focused on detecting hallucinations in summarization tasks in a RAG context. So the model is trained on that specifically as opposed to detecting &quot;arbitrary untruths&quot; in the output.<p>We do have an even more powerful model deployed in our platform, but even so, this is far better than anything else in the OSS realm today

暂无评论

暂无评论