搜索“Hugging Face”

cord) [Selected fields from the original training-selection record] This incident was identified through an ad hoc misalignment monitoring investigation by Jun 25, 2026. After the Hugging Face incident, retrospective misalignment monitoring of earlier reinforcement-learning and evaluation traffic flagged this incident as critical. In response to these and other incidents, we strengthened security controls, took further steps to prevent models from accessing the live internet during training, imp

匹配字段正文

4 [6]) is a dataset of 1 million conversations taken from between April 2023 and May 2024. It was collected by offering users free access to ChatGPT-3.5 & GPT-4 services hosted on Hugging Face Spaces, with an explicit opt-in consent to collect and release their de-identified chat transcripts. WildChat has since been used to construct several benchmarks [7,8,9] that attempt to rank AI models in terms of their capabilities or safety, by curating small sets of difficult prompts targeted at elicitin

匹配字段正文

gent’s ability to improve an open-source base model on a target benchmark. For each task, the agent receives a pretrained base model, one H100 GPU, internet access, a prepopulated Hugging Face dataset cache, and five hours. Within this budget, the agent must decide how to construct training data, which training method to apply, how to configure the training run, and how to use intermediate evaluation results to guide further experiments. The evaluation covers 12 combinations from the full PostTr

匹配字段正文

gent’s ability to improve an open-source base model on a target benchmark. For each task, the agent receives a pretrained base model, one H100 GPU, internet access, a prepopulated Hugging Face dataset cache, and five hours. Within this budget, the agent must decide how to construct training data, which training method to apply, how to configure the training run, and how to use intermediate evaluation results to guide further experiments. The evaluation covers 12 combinations from the full PostTr

匹配字段正文

model class. Extending it to Llama, Mistral, or MoE models requires no algorithmic changes. We hope these tools enable the community to explore architectures beyond what we've tested and to develop new algorithms and training recipes for building scalable next-generation architectures for AI agents. For full details, see our paper, "Priming: Hybrid state space models from pre-trained Transformers". Code, models, and serving plugin are available on GitHub, and the models are also on Hugging Face.

匹配字段正文

cking, and built probes that detect it — enabling efficient, real-time detection of reward hacking at scale. In July, a collective of hundreds of OpenAI agents autonomously hacked Hugging Face. Unlike most hackers, they were not after money, blackmail, or intellectual property. Instead, they were doing reconnaissance to figure out how to get away with cheating on an evaluation. This was an unusually stark example of reward hacking, a growing problem with agentic AI models. AI agents are like amo

匹配字段正文

ipts aren’t modified. But misaligned AI systems with superhuman hacking skills could subvert many or all of the systems companies rely on to detect AI misbehavior. In the OpenAI / Hugging Face incident, for example, models severely compromised a chunk of OpenAI’s internal infrastructure. If this incident happened with more capable and/or more misaligned models, they might cover up most or all of the evidence, hiding it from employees at the company. Speculatively in the worst case, AIs could cre

匹配字段正文

circle and, like, make projects with it would, immediately get, go viral. so like Swyx [00:16:55]: And you didn’t know who they are, right? They’re just some name on a, GitHub or Hugging Face somewhere. Anastasis [00:16:59]: Exactly, yeah. So I remember one of the first big viral moments of creative AI was, there was the neural style transfer paper Swyx [00:17:09]: Huh Anastasis [00:17:09]: That Swyx [00:17:10]: Something dreaming? Anastasis [00:17:11]: I think it was called neural style transf

匹配字段正文

ance to be featured. 🔀 Switch - Drop any AI agent into your Slack, Teams, or Discord and put it to work. Runs local. Get it on GitHub ⚙️ Halo - Open-source framework for 3× faster Hugging Face model training 📽️ Dramagic - ByteDance’s AIGC platform for short dramas and video production 🩺 Scribe v2 Medical - ElevenLabs’ speech-to-text model for medical transcription Sponsored listing Australian PM Anthony Albanese revealed that an OpenAI agent broke into an Australian Medicare statistics portal in

匹配字段正文

ing, narrowing, or excluding tasks when we found a concrete problem. Reported evaluations are closed-book. We block the public web and upstream source repositories, while allowing Hugging Face access for model weights, because an open-network pilot showed models retrieving task-specific upstream code. We audited all 1,749 trials behind the leaderboard; 196 prohibited retrieval attempts were blocked, and none succeeded. The paper describes the full qualification and evaluation-integrity process.

匹配字段正文

ier, but showing strong results in agentic, coding, and cyber benchmarks. Pricing comes in at just $0.15 / 0.60 per million tokens, with DeepSeek publishing the model’s weights on Hugging Face with an MIT License for download. Flash joins Z.ai ’s GLM-5.3-Flash as the two strongest intelligence systems in their price range, with DeepSeek still set to release larger models in the new family. Why it matters: DeepSeek's efficiency and pricing remain its edge, with Chinese labs in general (perhaps he

匹配字段正文

es outside the stated scope, tasks that cannot be completed without probing or breaking the sandbox boundary, and a situation inspired by parts of the attack chain from the OpenAI-Hugging Face incident. As in the real incidents, Claude Mythos 5 is the most willing of the models we tested to act against out-of-scope parties or leave its sandbox. For example, it was willing to commit a severely harmful action in the CTF replication roughly 80% of the time. In contrast, Claude Opus 5 and Claude Myt

匹配字段正文

hrough some of the mechanics and implications. It is an impressive account and well worth reading. Below we’ll give our take on a summary of the incident just the same. Here’s the Hugging Face timeline: Beginning on July 9, AI agents exploit Hugging Face’s datasets API, which runs on Kubernetes. The agents write up a malicious README that gets the worker that usually splits hdf5 datasets into train/test to read instead from /proc/self/environ, and return all creds and source on that worker; The

匹配字段正文

: B200 and B300 from 8-GPU HGX systems, Rubin from the HGX Rubin NVL8 table. For RL systems, the precision contract spans: SGLang rollout; Megatron and TransformerEngine training; Hugging Face checkpoint conversion; Megatron-to-Hugging Face live weight export; Fine-grained high-precision exceptions Format Background MXFP8 MXFP8 is a microscaling FP8 format. TransformerEngine's MXFP8 documentation describes it as a Blackwell-native blockwise scaling recipe: every 32 consecutive E4M3 values share

匹配字段正文

es the same sequences, computes the policy update, and trains the actor; Miles then transfers the updated weights back to the live SGLang workers. The current FP8 path uses an FP8 Hugging Face checkpoint for rollout and a BF16 Megatron torch dist checkpoint for actor training. The two engines therefore represent one policy in different execution formats, making conversion, rescoring, and online update behavior part of the correctness boundary. Figure 2. Prompts flow to SGLang rollout; trajectori

匹配字段正文

that supports all quantization and sharding configurations Several new interfaces are needed on the SGLang side: Exposing model parallelism for replica creation: PR 20907; Mapping hugging face tensor with its corresponding SGLang tensor shard: PR 17326; A post-process weight engine call for GPU-local processing such as post-quantization similar to PR 15245 Which are merged in the miles targeted sglang-miles branch. During the weight update, the caller side operates as below: Initialization Durin

匹配字段正文

its built-in Marlin kernel as the INT4 inference backend. However, in practice we encountered a notable “format gap”: QAT training outputs weights in standard formats (similar to Hugging Face), while SGLang’s Marlin kernel requires weights to be specially packed and permuted so that the kernel can read them efficiently. Given that RL training requires frequent weight updates, we must solve format compatibility. We therefore designed a reverse restore weights before loading safety mechanism. Usi

匹配字段正文

In conclusion, NeMo Curator is a powerful tool that allows us to not only to scale our data processing capabilities, but optimize costs. Download the Zyda-2 dataset directly from Hugging Face and train higher-accuracy models. For more information see our GitHub tutorial. Appendix Analysis of Global Duplicates Figure 7: Distribution of cluster sizes of duplicates in global dataset (log-log scale). Figure 8: Distribution of cluster sizes of duplicates in DCLM (log-log scale). Figure 9: Distributi

匹配字段正文