We introduce DataComp for Language Models (DCLM), a testbed for controlled dataset experiments with the goal of improving language models. As part of DCLM, we provide a standardized corpus of 240T tokens extracted from Common Crawl, effective pretraining recipes based on the OpenLM framework, and a broad suite of 53 downstream evaluations. Participants in the DCLM benchmark can experiment with data curation strategies such as deduplication, filtering, and data mixing at model scales ranging from 412M to 7B parameters. As a baseline for DCLM, we conduct extensive experiments and find that model-based filtering is key to assembling a high-quality training set. The resulting dataset, DCLM-Baseline enables training a 7B parameter language model from scratch to 64% 5-shot accuracy on MMLU with 2.6T training tokens. Compared to MAP-Neo, the previous state-of-the-art in open-data language models, DCLM-Baseline represents a 6.6 percentage point improvement on MMLU while being trained with 40% less compute. Our baseline model is also comparable to Mistral-7B-v0.3 and Llama 3 8B on MMLU (63% & 66%), and performs similarly on an average of 53 natural language understanding tasks while being trained with 6.6x less compute than Llama 3 8B. Our results highlight the importance of dataset design for training language models and offer a starting point for further research on data curation.

Streaming feature selection techniques have become essential in processing real-time data streams, as they facilitate the identification of the most relevant attributes from continuously updating information. Despite their performance, current algorithms to streaming feature selection frequently fall short in managing biases and avoiding discrimination that could be perpetuated by sensitive attributes, potentially leading to unfair outcomes in the resulting models. To address this issue, we propose FairSFS, a novel algorithm for Fair Streaming Feature Selection, to uphold fairness in the feature selection process without compromising the ability to handle data in an online manner. FairSFS adapts to incoming feature vectors by dynamically adjusting the feature set and discerns the correlations between classification attributes and sensitive attributes from this revised set, thereby forestalling the propagation of sensitive data. Empirical evaluations show that FairSFS not only maintains accuracy that is on par with leading streaming feature selection methods and existing fair feature techniques but also significantly improves fairness metrics.

Privacy-preserving synthetic data has been increasingly adopted to share data within and across organizations while reducing privacy risks. The intuition is to train a generative model on the real data, draw samples from the model, and create new (synthetic) data points. As the original data may contain sensitive and/or personal information, synthetic data can be vulnerable to membership/property inference, reconstruction attacks, etc. [6, 13, 25, 29, 30, 57]. Thus, models should be trained to satisfy robust definitions like Differential Privacy (DP) [19, 20], which bounds the privacy leakage from the synthetic data. Combining generative models with DP has been advocated for or deployed by government agencies [2, 31, 46, 62], regulatory bodies [60, 61], and non-profit organizations [48, 63].

Obstructive sleep apnea (OSA) is frequent and responsible for cardiovascular complications and excessive daytime sleepiness. It is underdiagnosed due to the difficulty to access the gold standard for diagnosis, polysomnography (PSG). Alternative methods using smartphone sensors could be useful to increase diagnosis. The objective is to assess the performances of Apneal, an application that records the sound using a smartphone's microphone and movements thanks to a smartphone's accelerometer and gyroscope, to estimate patients' AHI. In this article, we perform a monocentric proof-of-concept study with a first manual scoring step, and then an automatic detection of respiratory events from the recorded signals using a sequential deep-learning model which was released internally at Apneal at the end of 2022 (version 0.1 of Apneal automatic scoring of respiratory events), in adult patients during in-hospital polysomnography.46 patients (women 34 per cent, mean BMI 28.7 kg per m2) were included. For AHI superior to 15, sensitivity of manual scoring was 0.91, and positive predictive value (PPV) 0.89. For AHI superior to 30, sensitivity was 0.85, PPV 0.94. We obtained an AUC-ROC of 0.85 and an AUC-PR of 0.94 for the identification of AHI superior to 15, and AUC-ROC of 0.95 and AUC-PR of 0.93 for AHI superior to 30. Promising results are obtained for the automatic annotations of events.This article shows that manual scoring of smartphone-based signals is possible and accurate compared to PSG-based scorings. Automatic scoring method based on a deep learning model provides promising results. A larger multicentric validation study, involving subjects with different SAHS severity is required to confirm these results.

In sequential event prediction, which finds applications in finance, retail, social networks, and healthcare, a crucial task is forecasting multiple future events within a specified time horizon. Traditionally, this has been addressed through autoregressive generation using next-event prediction models, such as Marked Temporal Point Processes. However, autoregressive methods use their own output for future predictions, potentially reducing quality as the prediction horizon extends. In this paper, we challenge traditional approaches by introducing a novel benchmark, HoTPP, specifically designed to evaluate a model's ability to predict event sequences over a horizon. This benchmark features a new metric inspired by object detection in computer vision, addressing the limitations of existing metrics in assessing models with imprecise time-step predictions. Our evaluations on established datasets employing various models demonstrate that high accuracy in next-event prediction does not necessarily translate to superior horizon prediction, and vice versa. HoTPP aims to serve as a valuable tool for developing more robust event sequence prediction methods, ultimately paving the way for further advancements in the field.

In this work, we propose a two-phased approach for real-time detection and deterrence of ransomware. To achieve this, we leverage the capabilities of eBPF (Extended Berkeley Packet Filter) and artificial intelligence to develop both proactive and reactive methods. In the first phase, we utilize signature based detection, where we employ custom eBPF programs to trace the execution of new processes and perform hash-based analysis against a known ransomware dataset. In the second, we employ a behavior-based technique that focuses on monitoring the process activities using a custom eBPF program and the creation of ransom notes, a prominent indicator of ransomware activity through the use of Natural Language Processing (NLP). By leveraging low-level tracing capabilities of eBPF and integrating NLP based machine learning algorithms, our solution achieves an impressive 99.76% accuracy in identifying ransomware incidents within a few seconds on the onset of zero-day attacks.

We use (multi)modal deep neural networks (DNNs) to probe for sites of multimodal integration in the human brain by predicting stereoencephalography (SEEG) recordings taken while human subjects watched movies. We operationalize sites of multimodal integration as regions where a multimodal vision-language model predicts recordings better than unimodal language, unimodal vision, or linearly-integrated language-vision models. Our target DNN models span different architectures (e.g., convolutional networks and transformers) and multimodal training techniques (e.g., cross-attention and contrastive learning). As a key enabling step, we first demonstrate that trained vision and language models systematically outperform their randomly initialized counterparts in their ability to predict SEEG signals. We then compare unimodal and multimodal models against one another. Because our target DNN models often have different architectures, number of parameters, and training sets (possibly obscuring those differences attributable to integration), we carry out a controlled comparison of two models (SLIP and SimCLR), which keep all of these attributes the same aside from input modality. Using this approach, we identify a sizable number of neural sites (on average 141 out of 1090 total sites or 12.94%) and brain regions where multimodal integration seems to occur. Additionally, we find that among the variants of multimodal training techniques we assess, CLIP-style training is the best suited for downstream prediction of the neural activity in these sites.

Large Language Models (LLMs) are routinely used in retrieval-augmented applications to orchestrate tasks and process inputs from users and other sources. These inputs, even in a single LLM interaction, can come from a variety of sources, of varying trustworthiness and provenance. This opens the door to prompt injection attacks, where the LLM receives and acts upon instructions from supposedly data-only sources, thus deviating from the user's original instructions. We define this as task drift, and we propose to catch it by scanning and analyzing the LLM's activations. We compare the LLM's activations before and after processing the external input in order to detect whether this input caused instruction drift. We develop two probing methods and find that simply using a linear classifier can detect drift with near perfect ROC AUC on an out-of-distribution test set. We show that this approach generalizes surprisingly well to unseen task domains, such as prompt injections, jailbreaks, and malicious instructions, without being trained on any of these attacks. Our setup does not require any modification of the LLM (e.g., fine-tuning) or any text generation, thus maximizing deployability and cost efficiency and avoiding reliance on unreliable model output. To foster future research on activation-based task inspection, decoding, and interpretability, we will release our large-scale TaskTracker toolkit, comprising a dataset of over 500K instances, representations from 4 SoTA language models, and inspection tools.

This work addresses limitations in recent 3D tracking-by-detection methods, focusing on identifying legitimate trajectories and addressing state estimation drift in Kalman filters. Current methods rely heavily on threshold-based filtering of false positive detections using detection scores to prevent ghost trajectories. However, this approach is inadequate for distant and partially occluded objects, where detection scores tend to drop, potentially leading to false positives exceeding the threshold. Additionally, the literature generally treats detections as precise localizations of objects. Our research reveals that noise in detections impacts localization information, causing trajectory drift for occluded objects and hindering recovery. To this end, we propose a novel online track validity mechanism that temporally distinguishes between legitimate and ghost tracks, along with a multi-stage observational gating process for incoming observations. This mechanism significantly improves tracking performance, with a $6.28\%$ in HOTA and a $17.87\%$ increase in MOTA. We also introduce a refinement to the Kalman filter that enhances noise mitigation in trajectory drift, leading to more robust state estimation for occluded objects. Our framework, RobMOT, outperforms state-of-the-art methods, including deep learning approaches, across various detectors, achieving up to a $4\%$ margin in HOTA and $6\%$ in MOTA. RobMOT excels under challenging conditions, such as prolonged occlusions and tracking distant objects, with up to a 59\% improvement in processing latency.

Kolmogorov-Arnold Networks (KANs) have very recently been introduced into the world of machine learning, quickly capturing the attention of the entire community. However, KANs have mostly been tested for approximating complex functions or processing synthetic data, while a test on real-world tabular datasets is currently lacking. In this paper, we present a benchmarking study comparing KANs and Multi-Layer Perceptrons (MLPs) on tabular datasets. The study evaluates task performance and training times. From the results obtained on the various datasets, KANs demonstrate superior or comparable accuracy and F1 scores, excelling particularly in datasets with numerous instances, suggesting robust handling of complex data. We also highlight that this performance improvement of KANs comes with a higher computational cost when compared to MLPs of comparable sizes.