Zhibin Gou

🚀 DeepSeek-R1-0528 is here! 🔹 Improved benchmark performance 🔹 Enhanced front-end capabilities 🔹 Reduced hallucinations 🔹 Supports JSON output & function calling ✅ Try it now: chat.deepseek.com/ 🔌 No change to API usage — docs here: api-docs.deepseek.com/guides… 🔗 Open-source weights: huggingface.co/deepseek-ai/D…

🚀 Introducing DeepSeek-V3! Biggest leap forward yet: ⚡ 60 tokens/second (3x faster than V2!) 💪 Enhanced capabilities 🛠 API compatibility intact 🌍 Fully open-source models & papers 🐋 1/n

Realization: the old style of “hallucinations research” via self-calibration is probably going to die down. I used to be very excited about it but now I am skeptical because giving models internet access (e.g., searchGPT, perplexity) is turning out to be way higher ROI. When ChatGPT first came out, it was clear that hallucinations was one of the most important issues of LLMs. In 2023 John Schulman gave a talk on (what I think is) a really good approach to reduce hallucinations. The idea was that you could turn some amount of hallucinations into "I don't know", and the model would be able to learn that boundary using RL. I expected people to implement that and language models would get really good at knowing when to answer and when to say I don't know. I still think the idea is very correct but empirically it doesn't seem like anyone made great progress on that in the past two years. One reason is that it is pretty hard to move models to be stricty more pareto-optimal on precision-recall. Often you would just be moving along the precision-recall boundary, and in practice people just care a lot more about recall; models that say "I don't know" too much are annoying. The much better thing than self-calibration is improve factuality by allowing language models to browse the internet. In the past, language models would hallucinate easily on queries like “what papers Barret Zoph wrote” but now with browsing they can answer those easy factual questions pretty well. It turns out that this type of research---giving models access to the internet---is way higher ROI than teaching language models to be more calibrated. In other words, let's say you are at 50% recall on some factual dataset. One approach would be to spend months of hallucinations research to improve calibration, and maybe get 45% recall with much lower hallucination rate. However, with a similar amount of research work you could probably augment the model with internet access and get much better precision and recall (e.g., 90%). Even though the two approaches are technically orthogonal, once you starting doing the latter, the former is reduced to only a second-order effect.

Excited to share our paper "Not All Tokens Are What You Need for Pretraining" received the #NeurIPS2024 Best Paper Runner-up award. I’ll be attending the conference from Wed to Sat. Feel free to reach out if you'd like to connect or attend our talk. openreview.net/forum?id=0NMz…

🚀 DeepSeek-R1-Lite-Preview is now live: unleashing supercharged reasoning power! 🔍 o1-preview-level performance on AIME & MATH benchmarks. 💡 Transparent thought process in real-time. 🛠️ Open-source models & API coming soon! 🌐 Try it now at chat.deepseek.com #DeepSeek

AI Math Olympiad winner is now Open Source! 🔥 > 7B model, scored 29/50 on the public and private Kaggle test sets. (Apache 2.0 licensed). > Base model: deepseek-math-7b-base > Two-stage fine-tuning to reach the leaderboard-winning solution. > Step 1: fine-tune large and diverse math datasets (templated with Chain of Thought prompting). > Step 2: fine-tune the Stage 1 checkpoint on a synthetic dataset of tool-integrated reasoning, where each math problem is decomposed into a sequence of rationales, Python programs, and their outputs. > Used MSFT's ToRA format for GPT4 outputs with code executions. > Integrated with Transformers. 🤗