Information about pretraining corpora used to train the current best-performing language models is seldom discussed: commercial models rarely detail their data, and even open models are often released without accompanying training data or recipes to reproduce them. As a result, it is challenging to conduct and advance scientific research on language modeling, such as understanding how training data impacts model capabilities and limitations. To facilitate scientific research on language model pretraining, we curate and release Dolma, a three-trillion-token English corpus, built from a diverse mixture of web content, scientific papers, code, public-domain books, social media, and encyclopedic materials. We extensively document Dolma, including its design principles, details about its construction, and a summary of its contents. We present analyses and experimental results on intermediate states of Dolma to share what we have learned about important data curation practices. Finally, we open-source our data curation toolkit to enable reproduction of our work as well as support further research in large-scale data curation.
论文检索
输入标题、作者或关键词，从 9,460 篇学术成果中精准定位
Communicative efficiency is a key topic in linguistics and cognitive psychology, with many studies demonstrating how the pressure to communicate with minimal effort guides the form of natural language. However, this phenomenon is rarely explored in signed languages. This paper shows how handshapes in American Sign Language (ASL) reflect these efficiency pressures and provides new evidence of communicative efficiency in the visual-gestural modality.We focus on hand configurations in native ASL signs and signs borrowed from English to compare efficiency pressures from both ASL and English usage. First, we develop new methodologies to quantify the articulatory effort needed to produce handshapes and the perceptual effort required to recognize them. Then, we analyze correlations between communicative effort and usage statistics in ASL or English. Our findings reveal that frequent ASL handshapes are easier to produce and that pressures for communicative efficiency mostly come from ASL usage, rather than from English lexical borrowing.
Zero-shot stance detection (ZSSD) aims to determine whether the author of a text is in favor, against, or neutral toward a target that is unseen during training. In this paper, we present EZ-STANCE, a large English ZSSD dataset with 47,316 annotated text-target pairs. In contrast to VAST, which is the only other large existing ZSSD dataset for English, EZ-STANCE is 2.5 times larger, includes both noun-phrase targets and claim targets that cover a wide range of domains, provides two challenging subtasks for ZSSD: target-based ZSSD and domain-based ZSSD, and contains much harder examples for the neutral class. We evaluate EZ-STANCE using state-of-the-art deep learning models. Furthermore, we propose to transform ZSSD into the NLI task by applying simple yet effective prompts to noun-phrase targets. Our experimental results show that EZ-STANCE is a challenging new benchmark, which provides significant research opportunities on English ZSSD. We publicly release our dataset and code at https://github.com/chenyez/EZ-STANCE.
Annotated datasets are an essential ingredient to train, evaluate, compare and productionalize supervised machine learning models. It is therefore imperative that annotations are of high quality. For their creation, good quality management and thereby reliable quality estimates are needed. Then, if quality is insufficient during the annotation process, rectifying measures can be taken to improve it. Quality estimation is often performed by having experts manually label instances as correct or incorrect. But checking all annotated instances tends to be expensive. Therefore, in practice, usually only subsets are inspected; sizes are chosen mostly without justification or regard to statistical power and more often than not, are relatively small. Basing estimates on small sample sizes, however, can lead to imprecise values for the error rate. Using unnecessarily large sample sizes costs money that could be better spent, for instance on more annotations. Therefore, we first describe in detail how to use confidence intervals for finding the minimal sample size needed to estimate the annotation error rate. Then, we propose applying acceptance sampling as an alternative to error rate estimation We show that acceptance sampling can reduce the required sample sizes up to 50% while providing the same statistical guarantees.
Quality-Aware Translation Models: Efficient Generation and Quality Estimation in a Single Model
PDF ↗Maximum-a-posteriori (MAP) decoding is the most widely used decoding strategy for neural machine translation (NMT) models. The underlying assumption is that model probability correlates well with human judgment, with better translations getting assigned a higher score by the model. However, research has shown that this assumption does not always hold, and generation quality can be improved by decoding to optimize a utility function backed by a metric or quality-estimation signal, as is done by Minimum Bayes Risk (MBR) or Quality-Aware decoding. The main disadvantage of these approaches is that they require an additional model to calculate the utility function during decoding, significantly increasing the computational cost. In this paper, we propose to make the NMT models themselves quality-aware by training them to estimate the quality of their own output. Using this approach for MBR decoding we can drastically reduce the size of the candidate list, resulting in a speed-up of two-orders of magnitude. When applying our method to MAP decoding we obtain quality gains similar or even superior to quality reranking approaches, but with the efficiency of single pass decoding.
With the growing complexity of fact verification tasks, the concern with “thoughtful” reasoning capabilities is increasing. However, recent fact verification benchmarks mainly focus on checking a narrow scope of semantic factoids within claims and lack an explicit logical reasoning process. In this paper, we introduce CHECKWHY, a challenging dataset tailored to a novel causal fact verification task: checking the truthfulness of the causal relation within claims through rigorous reasoning steps. CHECKWHY consists of over 19K “why” claim-evidence- argument structure triplets with supports, refutes, and not enough info labels. Each argument structure is composed of connected evidence, representing the reasoning process that begins with foundational evidence and progresses toward claim establishment. Through extensive experiments on state-of-the-art models, we validate the importance of incorporating the argument structure for causal fact verification. Moreover, the automated and human evaluation of argument structure generation reveals the difficulty in producing satisfying argument structure by fine-tuned models or Chain-of-Thought prompted LLMs, leaving considerable room for future improvements.
Understanding memorisation in language models has practical and societal implications, e.g., studying models’ training dynamics or preventing copyright infringements.Prior work defines memorisation as the causal effect of training with an instance on the model’s ability to predict that instance. This definition relies on a counterfactual: the ability to observe what would have happened had the model not seen that instance.Existing methods struggle to provide computationally efficient and accurate estimates of this counterfactual. Further, they often estimate memorisation for a model architecture rather than for a specific model instance. This paper fills an important gap in the literature, proposing a new, principled, and efficient method to estimate memorisation based on the difference-in-differences design from econometrics. Using this method, we characterise a model’s memorisation profile–its memorisation trends across training–by only observing its behaviour on a small set of instances throughout training.In experiments with the Pythia model suite, we find that memorisation (i) is stronger and more persistent in larger models, (ii) is determined by data order and learning rate, and (iii) has stable trends across model sizes, thus making memorisation in larger models predictable from smaller ones.
RomanSetu: Efficiently unlocking multilingual capabilities of Large Language Models via Romanization
PDF ↗This study addresses the challenge of extending Large Language Models (LLMs) to non-English languages, specifically those using non-Roman scripts. We propose an approach that utilizes the romanized form of text as an interface for LLMs, hypothesizing that its frequent informal use and shared tokens with English enhance cross-lingual alignment. Our approach involve the continual pretraining of a English LLM like Llama 2 on romanized text of non-English, non-Roman script languages, followed by instruction tuning on romanized data. The results indicate that romanized text not only reduces token fertility by 2x-4x but also matches if not outperforms native script representation across various NLU, NLG and MT tasks. Moreover, the embeddings computed on romanized text exhibit closer alignment with their English translations than those from the native script. Our approach presents a promising direction for leveraging the power of English LLMs in languages traditionally underrepresented in NLP research.
M4LE: A Multi-Ability Multi-Range Multi-Task Multi-Domain Long-Context Evaluation Benchmark for Large Language Models
PDF ↗Managing long sequences has become an important and necessary feature for large language models (LLMs). However, assessing their ability to handle long contexts remains a challenge. This paper introduces M^4LE, a \textbf{M}ulti-ability, \textbf{M}ulti-range, \textbf{M}ulti-task, \textbf{M}ulti-domain benchmark for \textbf{L}ong-context \textbf{E}valuation. It encompasses 36 NLP datasets, covering 11 types of tasks and 12 domains, providing a comprehensive test bed. To address the lack of tasks featuring naturally long sequences, we propose an automatic approach to convert short-sequence tasks into long-sequence scenarios. These scenarios evaluate LLMs’ long-context understanding across five key abilities: understanding of single or multiple relevant spans in long contexts based on explicit or semantic hints, and global context understanding. This automatic approach allows us to create instances evenly distributed from 1k to 8k input length. Our evaluation of 11 prominent LLMs reveals that 1) Current LLMs struggle to understand long context, particularly when tasks require multiple-span attention. 2) Semantic retrieval is more difficult for competent LLMs. 3) Models fine-tuned on longer text with position interpolation have comparable performance to those using Neural Tangent Kernel (NTK) aware scaling methods without fine-tuning. We make our benchmark publicly available to encourage future research in this challenging area.
Originating from China’s Shang Dynasty approximately 3,000 years ago, the Oracle Bone Script (OBS) is a cornerstone in the annals of linguistic history, predating many established writing systems. Despite the discovery of thousands of inscriptions, a vast expanse of OBS remains undeciphered, casting a veil of mystery over this ancient language. The emergence of modern AI technologies presents a novel frontier for OBS decipherment, challenging traditional NLP methods that rely heavily on large textual corpora, a luxury not afforded by historical languages. This paper introduces a novel approach by adopting image generation techniques, specifically through the development of Oracle Bone Script Decipher (OBSD). Utilizing a conditional diffusion-based strategy, OBSD generates vital clues for decipherment, charting a new course for AI-assisted analysis of ancient languages. To validate its efficacy, extensive experiments were conducted on an oracle bone script dataset, with quantitative results demonstrating the effectiveness of OBSD.
With the rapid development of Large Language Models (LLMs), increasing attention has been paid to their safety concerns. Consequently, evaluating the safety of LLMs has become an essential task for facilitating the broad applications of LLMs. Nevertheless, the absence of comprehensive safety evaluation benchmarks poses a significant impediment to effectively assess and enhance the safety of LLMs. In this work, we present SafetyBench, a comprehensive benchmark for evaluating the safety of LLMs, which comprises 11,435 diverse multiple choice questions spanning across 7 distinct categories of safety concerns. Notably, SafetyBench also incorporates both Chinese and English data, facilitating the evaluation in both languages. Our extensive tests over 25 popular Chinese and English LLMs in both zero-shot and few-shot settings reveal a substantial performance advantage for GPT-4 over its counterparts, and there is still significant room for improving the safety of current LLMs. We also demonstrate that the measured safety understanding abilities in SafetyBench are correlated with safety generation abilities. Data and evaluation guidelines are available at https://github.com/thu-coai/SafetyBench. Submission entrance and leaderboard are available at https://llmbench.ai/safety.
The advent of artificial intelligence has led to a growing emphasis on data-driven modeling in macroeconomics, with agent-based modeling (ABM) emerging as a prominent bottom-up simulation paradigm. In ABM, agents (*e.g.*, households, firms) interact within a macroeconomic environment, collectively generating market dynamics. Existing agent modeling typically employs predetermined rules or learning-based neural networks for decision-making. However, customizing each agent presents significant challenges, complicating the modeling of agent heterogeneity. Additionally, the influence of multi-period market dynamics and multifaceted macroeconomic factors are often overlooked in decision-making processes.In this work, we introduce **EconAgent**, a large language model-empowered agent with human-like characteristics for macroeconomic simulation. We first construct a simulation environment that incorporates various market dynamics driven by agents’ decisions regarding work and consumption. Through the perception module, we create heterogeneous agents with distinct decision-making mechanisms. Furthermore, we model the impact of macroeconomic trends using a memory module, which allows agents to reflect on past individual experiences and market dynamics.Simulation experiments show that EconAgent can make realistic decisions, leading to more reasonable macroeconomic phenomena compared to existing rule-based or learning-based agents. Our codes are released at https://github.com/tsinghua-fib-lab/ACL24-EconAgent.
We introduce Contrastive Activation Addition (CAA), a method for steering language models by modifying their activations during forward passes. CAA computes “steering vectors” by averaging the difference in residual stream activations between pairs of positive and negative examples of a particular behavior, such as factual versus hallucinatory responses. During inference, these steering vectors are added at all token positions after the user’s prompt with either a positive or negative coefficient, allowing precise control over the degree of the targeted behavior. We evaluate CAA’s effectiveness on Llama 2 Chat using multiple-choice behavioral question datasets and open-ended generation tasks. We demonstrate that CAA significantly alters model behavior, is effective over and on top of traditional methods like finetuning and system prompt design, and minimally reduces capabilities. Moreover, we gain deeper insights into CAA’s mechanisms by employing various activation space interpretation methods. CAA accurately steers model outputs and sheds light on how high-level concepts are represented in Large Language Models (LLMs).
Language Complexity and Speech Recognition Accuracy: Orthographic Complexity Hurts, Phonological Complexity Doesn’t
PDF ↗We investigate what linguistic factors affect the performance of Automatic Speech Recognition (ASR) models. We hypothesize that orthographic and phonological complexities both degrade accuracy. To examine this, we fine-tune the multilingual self-supervised pretrained model Wav2Vec2-XLSR-53 on 25 languages with 15 writing systems, and we compare their ASR accuracy, number of graphemes, unigram grapheme entropy, logographicity (how much word/morpheme-level information is encoded in the writing system), and number of phonemes. The results demonstrate that a high logographicity correlates with low ASR accuracy, while phonological complexity has no significant effect.
Recent studies show that large language models (LLMs) improve their performance through self-feedback on certain tasks while degrade on others. We discovered that such a contrary is due to LLM’s bias in evaluating their own output. In this paper, we formally define LLM’s self-bias – the tendency to favor its own generation – using two statistics. We analyze six LLMs (GPT-4, GPT-3.5, Gemini, LLaMA2, Mixtral and DeepSeek) on translation, constrained text generation, and mathematical reasoning tasks. We find that self-bias is prevalent in all examined LLMs across multiple languages and tasks. Our analysis reveals that while the self-refine pipeline improves the fluency and understandability of model outputs, it further amplifies self-bias. To mitigate such biases, we discover that larger model size and external feedback with accurate assessment can significantly reduce bias in the self-refine pipeline, leading to actual performance improvement in downstream tasks. The code and data are released at https://github.com/xu1998hz/llm_self_bias.
Despite the advancements in in-context learning (ICL) for large language models (LLMs), current research centers on specific prompt engineering, such as demonstration selection, with the expectation that a single iteration of demonstrations processing can generalize effectively to a given test sample. However, this perspective overlooks the potential benefits derived from multiple iterations involving demonstrations, a practice aligning more closely with the iterative decision-making process exhibited by humans, who often learn through analogy. In this study, we introduce a novel two-stage framework to boost ICL in LLMs. Specifically, our framework delineates the ICL process into two distinct stages: Deep-Thinking and test stages. The Deep-Thinking stage incorporates a unique attention mechanism, i.e., iterative enhanced attention, which enables multiple rounds of information accumulation. This mechanism operates by manipulating the Key-Value matrices without training, fostering enhanced understanding capabilities in LLMs by thinking demonstrations multiple times. We evaluated Deep-Thinking across a range of benchmarks and LLMs, showing its superior performance over vanilla ICL methods and its effectiveness in challenging tasks where demonstration selection is infeasible.
As large language models (LLMs) are increasingly deployed in user-facing applications, building trust and maintaining safety by accurately quantifying a model’s confidence in its prediction becomes even more important. However, finding effective ways to calibrate LLMs—especially when the only interface to the models is their generated text—remains a challenge. We propose APRICOT (Auxiliary prediction of confidence targets): A method to set confidence targets and train an additional model that predicts an LLM’s confidence based on its textual input and output alone. This approach has several advantages: It is conceptually simple, does not require access to the target model beyond its output, does not interfere with the language generation, and has a multitude of potential usages, for instance by verbalizing the predicted confidence or using it to re-prompting the LLM to accurately reflecting its uncertainty. We show how our approach performs competitively in terms of calibration error for white-box and black-box LLMs on closed-book question-answering to detect incorrect LLM answers.
SPZ: A Semantic Perturbation-based Data Augmentation Method with Zonal-Mixing for Alzheimer’s Disease Detection
PDF ↗Alzheimer’s Disease (AD), characterized by significant cognitive and functional impairment, necessitates the development of early detection techniques. Traditional diagnostic practices, such as cognitive assessments and biomarker analysis, are often invasive and costly. Deep learning-based approaches for non-invasive AD detection have been explored in recent studies, but the lack of accessible data hinders further improvements in detection performance. To address these challenges, we propose a novel semantic perturbation-based data augmentation method that essentially differs from existing techniques, which primarily rely on explicit data engineering. Our approach generates controlled semantic perturbations to enhance textual representations, aiding the model in identifying AD-specific linguistic patterns, particularly in scenarios with limited data availability. It learns contextual information and dynamically adjusts the perturbation degree for different linguistic features. This enhances the model’s sensitivity to AD-specific linguistic features and its robustness against natural language noise. Experimental results on the ADReSS challenge dataset demonstrate that our approach outperforms other strong and competitive deep learning methods.
Framing studies how individuals and societies make sense of the world, by communicating or representing complex issues through schema of interpretation. The framing of information in the mass media influences our interpretation of facts and corresponding decisions, so detecting and analysing it is essential to understand biases in the information we consume. Despite that, framing is still mostly examined manually, on a case-by-case basis, while existing large-scale automatic analyses using NLP methods are not mature enough to solve this task. In this survey we show that despite the growing interest to framing in NLP its current approaches do not capture those aspects which allow to frame, rather than simply convey, the message. To this end, we bring together definitions of frames and framing adopted in different disciplines; examine cognitive, linguistic, and communicative aspects a frame contains beyond its topical content. We survey recent work on computational frame detection, and discuss how framing aspects and frame definitions are (or should) be reflected in NLP approaches.
G-DIG: Towards Gradient-based DIverse and hiGh-quality Instruction Data Selection for Machine Translation
PDF ↗Large Language Models (LLMs) have demonstrated remarkable abilities in general scenarios. Instruction finetuning empowers them to align with humans in various tasks. Nevertheless, the Diversity and Quality of the instruction data remain two main challenges for instruction finetuning. With regard to this, in this paper, we propose a novel gradient-based method to automatically select high-quality and diverse instruction finetuning data for machine translation. Our key innovation centers around analyzing how individual training examples influence the model during training. Specifically, we select training examples that exert beneficial influences on the model as high-quality ones by means of Influence Function plus a small high-quality seed dataset. Moreover, to enhance the diversity of the training data we maximize the variety of influences they have on the model by clustering on their gradients and resampling. Extensive experiments on WMT22 and FLORES translation tasks demonstrate the superiority of our methods, and in-depth analysis further validates their effectiveness and generalization.