Significant advances have been made in natural language processing in recent years. However, our current deep learning approach to language modeling requires substantial resources in terms of data and computation. One of the side effects of this data-hungry paradigm is the current schism between languages, separating those considered high-resource, where most of the development happens and resources are available, and the low-resource ones, which struggle to attain the same level of performance and autonomy. This study aims to introduce a new set of resources to stimulate the future development of neural text generation in Portuguese. In this work, we document the development of GigaVerbo, a concatenation of deduplicated Portuguese text corpora amounting to 200 billion tokens. Via this corpus, we trained a series of decoder-transformers named Tucano. Our models perform equal or superior to other Portuguese and multilingual language models of similar size in several Portuguese benchmarks. The evaluation of our models also reveals that model performance on many currently available benchmarks used by the Portuguese NLP community has little to no correlation with the scaling of token ingestion during training, highlighting the limitations of such evaluations when it comes to the assessment of Portuguese generative language models. All derivatives of our study are openly released on GitHub and Hugging Face. See https://nkluge-correa.github.io/Tucano/
Recent work finds that retrieval-augmented generation with large language models is prone to be influenced by the order of retrieved documents in the context. However, the lack of in-depth analysis limits the use of this phenomenon for prompt engineering in practice. In this study, we posit that likelihoods serve as an effective gauge for language model performance. Through experiments on two question-answering datasets with a variety of state-of-the-art language models, we reveal correlations between answer accuracy and the likelihood of the question at both the corpus level and the instance level. In addition, we find that question likelihood can also indicate the position of the task-relevant information in the context. Based on these findings, we propose two methods that use question likelihood as a gauge for selecting and constructing prompts that lead to better performance. We demonstrate their effectiveness with experiments. In addition, our likelihood-based methods are efficient, as they only need to compute the likelihood of the input, requiring much fewer language model passes than heuristic prompt engineering methods that require generating responses. Our analysis deepens our understanding of how input prompts affect model performance and provides a promising direction for efficient prompt optimization.
Quantization stands as a pivotal technique for large language model (LLM) serving, yet it poses significant challenges particularly in achieving effective low-bit quantization. The limited numerical mapping makes the quantized model produce a non-trivial error, bringing out intolerable performance degration. This paper is anchored in the basic idea of model compression objectives, and delves into the layer-wise error distribution of LLMs during post-training quantization. Subsequently, we introduce ASER, an algorithm consisting of (1) Error Reconstruction: low-rank compensation for quantization error with LoRA-style matrices constructed by whitening SVD; (2) Activation Smoothing: outlier extraction to gain smooth activation and better error compensation. ASER is capable of quantizing typical LLMs to low-bit ones, particularly preserving accuracy even in W4A8 per-channel setup. Experimental results show that ASER is competitive among the state-of-the-art quantization algorithms, showing potential to activation quantization, with minor overhead.
In 2022, with the release of ChatGPT, large-scale language models gained widespread attention. ChatGPT not only surpassed previous models in terms of parameters and the scale of its pretraining corpus but also achieved revolutionary performance improvements through fine-tuning on a vast amount of high-quality, human-annotated data. This progress has led enterprises and research institutions to recognize that building smarter and more powerful models relies on rich and high-quality datasets. Consequently, the construction and optimization of datasets have become a critical focus in the field of artificial intelligence. This paper summarizes the current state of pretraining and fine-tuning data for training large-scale language models, covering aspects such as data scale, collection methods, data types and characteristics, processing workflows, and provides an overview of available open-source datasets.
With the proliferation of the Large Language Model (LLM), the concept of World Models (WM) has recently attracted a great deal of attention in the AI research community, especially in the context of AI agents. It is arguably evolving into an essential foundation for building AI agent systems. A WM is intended to help the agent predict the future evolution of environmental states or help the agent fill in missing information so that it can plan its actions and behave safely. The safety property of WM plays a key role in their effective use in critical applications. In this work, we review and analyze the impacts of the current state-of-the-art in WM technology from the point of view of trustworthiness and safety based on a comprehensive survey and the fields of application envisaged. We provide an in-depth analysis of state-of-the-art WMs and derive technical research challenges and their impact in order to call on the research community to collaborate on improving the safety and trustworthiness of WM.
A pre-trained visual-language model, contrastive language-image pre-training (CLIP), successfully accomplishes various downstream tasks with text prompts, such as finding images or localizing regions within the image. Despite CLIP's strong multi-modal data capabilities, it remains limited in specialized environments, such as medical applications. For this purpose, many CLIP variants-i.e., BioMedCLIP, and MedCLIP-SAMv2-have emerged, but false positives related to normal regions persist. Thus, we aim to present a simple yet important goal of reducing false positives in medical anomaly detection. We introduce a Contrastive LAnguage Prompting (CLAP) method that leverages both positive and negative text prompts. This straightforward approach identifies potential lesion regions by visual attention to the positive prompts in the given image. To reduce false positives, we attenuate attention on normal regions using negative prompts. Extensive experiments with the BMAD dataset, including six biomedical benchmarks, demonstrate that CLAP method enhances anomaly detection performance. Our future plans include developing an automated fine prompting method for more practical usage.
Model stealing, where a learner tries to recover an unknown model via carefully chosen queries, is a critical problem in machine learning, as it threatens the security of proprietary models and the privacy of data they are trained on. In recent years, there has been particular interest in stealing large language models (LLMs). In this paper, we aim to build a theoretical understanding of stealing language models by studying a simple and mathematically tractable setting. We study model stealing for Hidden Markov Models (HMMs), and more generally low-rank language models. We assume that the learner works in the conditional query model, introduced by Kakade, Krishnamurthy, Mahajan and Zhang. Our main result is an efficient algorithm in the conditional query model, for learning any low-rank distribution. In other words, our algorithm succeeds at stealing any language model whose output distribution is low-rank. This improves upon the previous result by Kakade, Krishnamurthy, Mahajan and Zhang, which also requires the unknown distribution to have high "fidelity", a property that holds only in restricted cases. There are two key insights behind our algorithm: First, we represent the conditional distributions at each timestep by constructing barycentric spanners among a collection of vectors of exponentially large dimension. Second, for sampling from our representation, we iteratively solve a sequence of convex optimization problems that involve projection in relative entropy to prevent compounding of errors over the length of the sequence. This is an interesting example where, at least theoretically, allowing a machine learning model to solve more complex problems at inference time can lead to drastic improvements in its performance.
Large and Small Language Models (LMs) are typically pretrained using extensive volumes of text, which are sourced from publicly accessible platforms such as Wikipedia, Book Corpus, or through web scraping. These models, due to their exposure to a wide range of language data, exhibit impressive generalization capabilities and can perform a multitude of tasks simultaneously. However, they often fall short when it comes to domain-specific tasks due to their broad training data. This paper introduces SecEncoder, a specialized small language model that is pretrained using security logs. SecEncoder is designed to address the domain-specific limitations of general LMs by focusing on the unique language and patterns found in security logs. Experimental results indicate that SecEncoder outperforms other LMs, such as BERTlarge, DeBERTa-v3-large and OpenAI's Embedding (textembedding-ada-002) models, which are pretrained mainly on natural language, across various tasks. Furthermore, although SecEncoder is primarily pretrained on log data, it outperforms models pretrained on natural language for a range of tasks beyond log analysis, such as incident prioritization and threat intelligence document retrieval. This suggests that domain specific pretraining with logs can significantly enhance the performance of LMs in security. These findings pave the way for future research into security-specific LMs and their potential applications.
This paper introduces TIPS: Threat Actor Informed Prioritization using SecEncoder, a specialized language model for security. TIPS combines the strengths of both encoder and decoder language models to detect and prioritize compromised applications. By integrating threat actor intelligence, TIPS enhances the accuracy and relevance of its detections. Extensive experiments with a real-world benchmark dataset of applications demonstrate TIPS's high efficacy, achieving an F-1 score of 0.90 in identifying malicious applications. Additionally, in real-world scenarios, TIPS significantly reduces the backlog of investigations for security analysts by 87%, thereby streamlining the threat response process and improving overall security posture.
Impersonation tactics, such as app squatting and app cloning, have posed longstanding challenges in mobile app stores, where malicious actors exploit the names and reputations of popular apps to deceive users. With the rapid growth of Large Language Model (LLM) stores like GPT Store and FlowGPT, these issues have similarly surfaced, threatening the integrity of the LLM app ecosystem. In this study, we present the first large-scale analysis of LLM app squatting and cloning using our custom-built tool, LLMappCrazy. LLMappCrazy covers 14 squatting generation techniques and integrates Levenshtein distance and BERT-based semantic analysis to detect cloning by analyzing app functional similarities. Using this tool, we generated variations of the top 1000 app names and found over 5,000 squatting apps in the dataset. Additionally, we observed 3,509 squatting apps and 9,575 cloning cases across six major platforms. After sampling, we find that 18.7% of the squatting apps and 4.9% of the cloning apps exhibited malicious behavior, including phishing, malware distribution, fake content dissemination, and aggressive ad injection.
Deceptive patterns (DPs) in digital interfaces manipulate users into making unintended decisions, exploiting cognitive biases and psychological vulnerabilities. These patterns have become ubiquitous across various digital platforms. While efforts to mitigate DPs have emerged from legal and technical perspectives, a significant gap in usable solutions that empower users to identify and make informed decisions about DPs in real-time remains. In this work, we introduce AutoBot, an automated, deceptive pattern detector that analyzes websites' visual appearances using machine learning techniques to identify and notify users of DPs in real-time. AutoBot employs a two-staged pipeline that processes website screenshots, identifying interactable elements and extracting textual features without relying on HTML structure. By leveraging a custom language model, AutoBot understands the context surrounding these elements to determine the presence of deceptive patterns. We implement AutoBot as a lightweight Chrome browser extension that performs all analyses locally, minimizing latency and preserving user privacy. Through extensive evaluation, we demonstrate AutoBot's effectiveness in enhancing users' ability to navigate digital environments safely while providing a valuable tool for regulators to assess and enforce compliance with DP regulations.
This paper explores a new black-box, zero-shot language model inversion problem and proposes an innovative framework for prompt reconstruction using only text outputs from a language model. Leveraging a large language model alongside an optimization algorithm, the proposed method effectively recovers prompts with minimal resources. Experimental results on several datasets derived from public sources indicate that the proposed approach achieves high-quality prompt recovery and generates prompts more similar to the originals than current state-of-the-art methods. Additionally, the use-case study demonstrates the method's strong potential for generating high-quality text data.
Contextual ASR or hotword customization holds substantial practical value. Despite the impressive performance of current end-to-end (E2E) automatic speech recognition (ASR) systems, they often face challenges in accurately recognizing rare words. Typical E2E contextual ASR models commonly feature complex architectures and decoding mechanisms, limited in performance and susceptible to interference from distractor words. With large language model (LLM)-based ASR models emerging as the new mainstream, we propose a CTC-Assisted LLM-Based Contextual ASR model with an efficient filtering algorithm. By using coarse CTC decoding results to filter potential relevant hotwords and incorporating them into LLM prompt input, our model attains WER/B-WER of 1.27%/3.67% and 2.72%/8.02% on the Librispeech test-clean and test-other sets targeting on recognizing rare long-tail words, demonstrating significant improvements compared to the baseline LLM-based ASR model, and substantially surpassing other related work. More remarkably, with the help of the large language model and proposed filtering algorithm, our contextual ASR model still performs well with 2000 biasing words.
In this technical report, we present Zyda-2: a five trillion token dataset for language model pretraining. Zyda-2 was used to train our Zamba2 series of models which are state-of-the-art for their weight class. We build Zyda-2 by collating high-quality open-source tokens such as FineWeb and DCLM, then distilling them to the highest-quality subset via cross-deduplication and model-based quality filtering. Zyda-2 is released under a permissive open license, and is available at https://huggingface.co/datasets/Zyphra/Zyda-2
Data sanitization in the context of language modeling involves identifying sensitive content, such as personally identifiable information (PII), and redacting them from a dataset corpus. It is a common practice used in natural language processing (NLP) to maintain privacy. Nevertheless, the impact of data sanitization on the language understanding capability of a language model remains less studied. This paper empirically analyzes the effects of data sanitization across several benchmark language-modeling tasks including comprehension question answering (Q&A), entailment, sentiment analysis, and text classification. Our experiments cover a wide spectrum comprising finetuning small-scale language models, to prompting large language models (LLMs), on both original and sanitized datasets, and comparing their performance across the tasks. Interestingly, our results suggest that for some tasks such as sentiment analysis or entailment, the impact of redaction is quite low, typically around 1-5%, while for tasks such as comprehension Q&A there is a big drop of >25% in performance observed in redacted queries as compared to the original. For tasks that have a higher impact, we perform a deeper dive to inspect the presence of task-critical entities. Finally, we investigate correlation between performance and number of redacted entities, and also suggest a strategy to repair an already redacted dataset by means of content-based subsampling. Additional details are available at https://sites.google.com/view/datasan.
OpenThaiGPT 1.5 is an advanced Thai language chat model based on Qwen v2.5, finetuned on over 2,000,000 Thai instruction pairs. This report provides an engineering perspective on the model's development, capabilities, and performance. We discuss the model's architecture, training process, and key features, including multi-turn conversation support, Retrieval Augmented Generation (RAG) compatibility, and tool-calling functionality. Benchmark results demonstrate OpenThaiGPT 1.5's state-of-the-art performance on various Thai language tasks, outperforming other open-source Thai language models. We also address practical considerations such as GPU memory requirements and deployment strategies.
Language model users often issue queries that lack specification, where the context under which a query was issued -- such as the user's identity, the query's intent, and the criteria for a response to be useful -- is not explicit. For instance, a good response to a subjective query like "What book should I read next?" would depend on the user's preferences, and a good response to an open-ended query like "How do antibiotics work against bacteria?" would depend on the user's expertise. This makes evaluation of responses to such queries an ill-posed task, as evaluators may make arbitrary judgments about the response quality. To remedy this, we present contextualized evaluations, a protocol that synthetically constructs context surrounding an underspecified query and provides it during evaluation. We find that the presence of context can 1) alter conclusions drawn from evaluation, even flipping win rates between model pairs, 2) nudge evaluators to make fewer judgments based on surface-level criteria, like style, and 3) provide new insights about model behavior across diverse contexts. Specifically, our procedure uncovers an implicit bias towards WEIRD contexts in models' "default" responses and we find that models are not equally sensitive to following different contexts, even when they are provided in prompts.
This technical report presents our initial attempt to build a spoken large language model (LLM) for Taiwanese Mandarin, specifically tailored to enable real-time, speech-to-speech interaction in multi-turn conversations. Our end-to-end model incorporates a decoder-only transformer architecture and aims to achieve seamless interaction while preserving the conversational flow, including full-duplex capabilities allowing simultaneous speaking and listening. The paper also details the training process, including data preparation with synthesized dialogues and adjustments for real-time interaction. We also developed a platform to evaluate conversational fluency and response coherence in multi-turn dialogues. We hope the release of the report can contribute to the future development of spoken LLMs in Taiwanese Mandarin.
In large language model (LLM) training, several parallelization strategies, including Tensor Parallelism (TP), Pipeline Parallelism (PP), Data Parallelism (DP), as well as Sequence Parallelism (SP) and Context Parallelism (CP), are employed to distribute model parameters, activations, and optimizer states across devices. Identifying the optimal parallelization configuration for each environment while avoiding GPU memory overflow remains a challenging task. In this study, we provide precise formulas to estimate the memory consumed by parameters, gradients, optimizer states, and activations for 4D parallel training (DP, TP, PP, CP) in the Llama architecture. We conducted 454 experiments on A100 and H100 GPUs, incorporating often neglected factors such as temporary buffers and memory fragmentation into our analysis. Results indicate that when the estimated memory usage is below 80\% of the available GPU memory, the training never encounters out-of-memory errors. This simple yet effective formula allows us to identify parallelization configurations that could lead to memory overflow in advance, significantly reducing the configuration search space. Additionally, through a comprehensive exploration of optimal configurations in 4D parallelism, our analysis of the 454 experimental results provides empirical insights into optimal 4D parallelism configurations.
This paper proposes a Clustering, Labeling, then Augmenting framework that significantly enhances performance in Semi-Supervised Text Classification (SSTC) tasks, effectively addressing the challenge of vast datasets with limited labeled examples. Unlike traditional SSTC approaches that rely on a predefined small set of labeled data to generate pseudo-labels for the unlabeled data, this framework innovatively employs clustering to select representative "landmarks" for labeling. These landmarks subsequently act as intermediaries in an ensemble of augmentation techniques, including Retrieval-Augmented Generation (RAG), Large Language Model (LLMs)-based rewriting, and synonym substitution, to generate synthetic labeled data without making pseudo-labels for the unlabeled data. Empirical results show that even in complex text document classification scenarios involving over 100 categories, our method achieves state-of-the-art accuracies of 95.41% on the Reuters dataset and 82.43% on the Web of Science dataset. Our approach significantly reduces the reliance on human labeling efforts and the associated expenses, while simultaneously ensuring high data quality and minimizing privacy risks. The finetuning results further show the efficiency of fine-tuning LLMs for text classification tasks, highlighting a robust solution for leveraging limited labeled data.