A mechanistic understanding of how MLPs do computation in deep neural networks remains elusive. Current interpretability work can extract features from hidden activations over an input dataset but generally cannot explain how MLP weights construct features. One challenge is that element-wise nonlinearities introduce higher-order interactions and make it difficult to trace computations through the MLP layer. In this paper, we analyze bilinear MLPs, a type of Gated Linear Unit (GLU) without any element-wise nonlinearity that nevertheless achieves competitive performance. Bilinear MLPs can be fully expressed in terms of linear operations using a third-order tensor, allowing flexible analysis of the weights. Analyzing the spectra of bilinear MLP weights using eigendecomposition reveals interpretable low-rank structure across toy tasks, image classification, and language modeling. We use this understanding to craft adversarial examples, uncover overfitting, and identify small language model circuits directly from the weights alone. Our results demonstrate that bilinear layers serve as an interpretable drop-in replacement for current activation functions and that weight-based interpretability is viable for understanding deep-learning models.
The performance of Large Language Models (LLMs) is substantially influenced by the pretraining corpus, which consists of vast quantities of unsupervised data processed by the models. Despite its critical role in model performance, ensuring the quality of this data is challenging due to its sheer volume and the absence of sample-level quality annotations and enhancements. In this paper, we introduce DecorateLM, a data engineering method designed to refine the pretraining corpus through data rating, tagging and editing. Specifically, DecorateLM rates texts against quality criteria, tags texts with hierarchical labels, and edits texts into a more formalized format. Due to the massive size of the pretraining corpus, adopting an LLM for decorating the entire corpus is less efficient. Therefore, to balance performance with efficiency, we curate a meticulously annotated training corpus for DecorateLM using a large language model and distill data engineering expertise into a compact 1.2 billion parameter small language model (SLM). We then apply DecorateLM to enhance 100 billion tokens of the training corpus, selecting 45 billion tokens that exemplify high quality and diversity for the further training of another 1.2 billion parameter LLM. Our results demonstrate that employing such high-quality data can significantly boost model performance, showcasing a powerful approach to enhance the quality of the pretraining corpus.
In contexts with limited computational and data resources, high-resource language models often prove inadequate, particularly when addressing the specific needs of Malay languages. This paper introduces a Personal Intelligence System designed to efficiently integrate both on-device and server-based models. The system incorporates SLiM-34M for on-device processing, optimized for low memory and power usage, and MANYAK-1.3B for server-based tasks, allowing for scalable, high-performance language processing. The models achieve significant results across various tasks, such as machine translation, question-answering, and translate IndoMMLU. Particularly noteworthy is SLiM-34M's ability to achieve a high improvement in accuracy compared to other LLMs while using 2 times fewer pre-training tokens. This work challenges the prevailing assumption that large-scale computational resources are necessary to build effective language models, contributing to the development of resource-efficient models for the Malay language with the unique orchestration between SLiM-34M and MANYAK-1.3B.
Automatic question generation (QG) serves a wide range of purposes, such as augmenting question-answering (QA) corpora, enhancing chatbot systems, and developing educational materials. Despite its importance, most existing datasets predominantly focus on English, resulting in a considerable gap in data availability for other languages. Cross-lingual transfer for QG (XLT-QG) addresses this limitation by allowing models trained on high-resource language datasets to generate questions in low-resource languages. In this paper, we propose a simple and efficient XLT-QG method that operates without the need for monolingual, parallel, or labeled data in the target language, utilizing a small language model. Our model, trained solely on English QA datasets, learns interrogative structures from a limited set of question exemplars, which are then applied to generate questions in the target language. Experimental results show that our method outperforms several XLT-QG baselines and achieves performance comparable to GPT-3.5-turbo across different languages. Additionally, the synthetic data generated by our model proves beneficial for training multilingual QA models. With significantly fewer parameters than large language models and without requiring additional training for target languages, our approach offers an effective solution for QG and QA tasks across various languages.
Generating domain-specific content using small language models poses challenges, especially when dealing with multiple distinct datasets with minimal overlap. In this study, we explore methods to enable a small language model to produce coherent and relevant outputs for two different domains: stories (Dataset A) and recipes (Dataset B). Our initial experiments show that training individual models on each dataset yields satisfactory results, with each model generating appropriate content within its domain. We find that utilizing custom tokenizers tailored to each dataset significantly enhances generation quality compared to using a generic tokenizer. Attempts to adapt a single model to both domains using Low-Rank Adaptation (LoRA) or standard fine-tuning do not yield substantial results, often failing to produce meaningful outputs. Moreover, full fine-tuning without freezing the model's existing weights leads to catastrophic forgetting, where the model loses previously learned information and only retains knowledge from the new data. To overcome these challenges, we employ a knowledge expansion strategy: training only with additional parameters. This approach enables the model to generate both stories and recipes upon request, effectively handling multiple domains without suffering from catastrophic forgetting. Our findings demonstrate that knowledge expansion with frozen layers is an effective method for small language models to generate domain-specific content across distinct datasets. This work contributes to the development of efficient multi-domain language models and provides insights into managing catastrophic forgetting in small-scale architectures.
Chain-of-Thought (CoT) reasoning has enabled Large Language Model (LLM) to achieve remarkable performance in various NLP tasks, including arithmetic problem-solving. However, this success does not generalize to small language model (sLM) like T5, due to their limited capacity and absence of emergent abilities associated with larger models. Recent works to enhance sLM through knowledge distillation have yielded some improvements but still face significant limitations, particularly high ambiguity from the variability in natural language expressions and substantial computational costs. In this paper, we investigate why sLM perform poorly on arithmetic reasoning tasks and hypothesize that natural language format variability introduces high ambiguity for these smaller models. Based on this hypothesis, we conduct experiments with equation-only format, which is a reasoning format that unifies arithmetic reasoning previously expressed in natural language formats into mathematical equations. Experiment results demonstrate that equation-only format effectively boosts the arithmetic reasoning abilities of sLM, especially in very small models like T5-Tiny.
In this paper, we evaluate the creative fiction writing abilities of a fine-tuned small language model (SLM), BART-large, and compare its performance to human writers and two large language models (LLMs): GPT-3.5 and GPT-4o. Our evaluation consists of two experiments: (i) a human study in which 68 participants rated short stories from humans and the SLM on grammaticality, relevance, creativity, and attractiveness, and (ii) a qualitative linguistic analysis examining the textual characteristics of stories produced by each model. In the first experiment, BART-large outscored average human writers overall (2.11 vs. 1.85), a 14% relative improvement, though the slight human advantage in creativity was not statistically significant. In the second experiment, qualitative analysis showed that while GPT-4o demonstrated near-perfect coherence and used less cliche phrases, it tended to produce more predictable language, with only 3% of its synopses featuring surprising associations (compared to 15% for BART). These findings highlight how model size and fine-tuning influence the balance between creativity, fluency, and coherence in creative writing tasks, and demonstrate that smaller models can, in certain contexts, rival both humans and larger models.
Entity matching (EM) is the problem of determining whether two records refer to same real-world entity, which is crucial in data integration, e.g., for product catalogs or address databases. A major drawback of many EM approaches is their dependence on labelled examples. We thus focus on the challenging setting of zero-shot entity matching where no labelled examples are available for an unseen target dataset. Recently, large language models (LLMs) have shown promising results for zero-shot EM, but their low throughput and high deployment cost limit their applicability and scalability. We revisit the zero-shot EM problem with AnyMatch, a small language model fine-tuned in a transfer learning setup. We propose several novel data selection techniques to generate fine-tuning data for our model, e.g., by selecting difficult pairs to match via an AutoML filter, by generating additional attribute-level examples, and by controlling label imbalance in the data. We conduct an extensive evaluation of the prediction quality and deployment cost of our model, in a comparison to thirteen baselines on nine benchmark datasets. We find that AnyMatch provides competitive prediction quality despite its small parameter size: it achieves the second-highest F1 score overall, and outperforms several other approaches that employ models with hundreds of billions of parameters. Furthermore, our approach exhibits major cost benefits: the average prediction quality of AnyMatch is within 4.4% of the state-of-the-art method MatchGPT with the proprietary trillion-parameter model GPT-4, yet AnyMatch requires four orders of magnitude less parameters and incurs a 3,899 times lower inference cost (in dollars per 1,000 tokens).
Recent large language models (LLMs) have enabled the development of advanced agentic systems that can integrate various tools and APIs to fulfill user queries through function calling. However, the deployment of these LLMs on the edge has not been explored since they typically require cloud-based infrastructure due to their substantial model size and computational demands. To this end, we present TinyAgent, an end-to-end framework for training and deploying task-specific small language model agents capable of function calling for driving agentic systems at the edge. We first show how to enable accurate function calling for open-source models via the LLMCompiler framework. We then systematically curate a high-quality dataset for function calling, which we use to fine-tune two small language models, TinyAgent-1.1B and 7B. For efficient inference, we introduce a novel tool retrieval method to reduce the input prompt length and utilize quantization to further accelerate the inference speed. As a driving application, we demonstrate a local Siri-like system for Apple's MacBook that can execute user commands through text or voice input. Our results show that our models can achieve, and even surpass, the function-calling capabilities of larger models like GPT-4-Turbo, while being fully deployed at the edge. We open-source our dataset, models, and installable package and provide a demo video for our MacBook assistant agent.
This paper examines the performance of two Large Language Models (LLMs), GPT3.5 and Llama2 and one Small Language Model (SLM) Gemma, across three different classification tasks within the climate change (CC) and environmental domain. Employing BERT-based models as a baseline, we compare their efficacy against these transformer-based models. Additionally, we assess the models' self-evaluation capabilities by analyzing the calibration of verbalized confidence scores in these text classification tasks. Our findings reveal that while BERT-based models generally outperform both the LLMs and SLM, the performance of the large generative models is still noteworthy. Furthermore, our calibration analysis reveals that although Gemma is well-calibrated in initial tasks, it thereafter produces inconsistent results; Llama is reasonably calibrated, and GPT consistently exhibits strong calibration. Through this research, we aim to contribute to the ongoing discussion on the utility and effectiveness of generative LMs in addressing some of the planet's most urgent issues, highlighting their strengths and limitations in the context of ecology and CC.
High-resource language models often fall short in the African context, where there is a critical need for models that are efficient, accessible, and locally relevant, even amidst significant computing and data constraints. This paper introduces InkubaLM, a small language model with 0.4 billion parameters, which achieves performance comparable to models with significantly larger parameter counts and more extensive training data on tasks such as machine translation, question-answering, AfriMMLU, and the AfriXnli task. Notably, InkubaLM outperforms many larger models in sentiment analysis and demonstrates remarkable consistency across multiple languages. This work represents a pivotal advancement in challenging the conventional paradigm that effective language models must rely on substantial resources. Our model and datasets are publicly available at https://huggingface.co/lelapa to encourage research and development on low-resource languages.
The widespread adoption of cloud-based proprietary large language models (LLMs) has introduced significant challenges, including operational dependencies, privacy concerns, and the necessity of continuous internet connectivity. In this work, we introduce an LLMOps pipeline, "LlamaDuo", for the seamless migration of knowledge and abilities from service-oriented LLMs to smaller, locally manageable models. This pipeline is crucial for ensuring service continuity in the presence of operational failures, strict privacy policies, or offline requirements. Our LlamaDuo involves fine-tuning a small language model against the service LLM using a synthetic dataset generated by the latter. If the performance of the fine-tuned model falls short of expectations, it is automatically improved through additional fine-tuning using extra similar data generated by the service LLM. This multi-turn process guarantees that the smaller model can eventually match or even surpass the service LLM's capabilities in specific downstream tasks, offering a practical and scalable solution for managing AI deployments in constrained environments. Extensive experiments with leading-edge LLMs are conducted to demonstrate the effectiveness, adaptability, and affordability of LlamaDuo across various downstream tasks. Our pipeline implementation is available at https://github.com/deep-diver/llamaduo.
Large language models (LLMs) are highly capable but face latency challenges in real-time applications, such as conducting online hallucination detection. To overcome this issue, we propose a novel framework that leverages a small language model (SLM) classifier for initial detection, followed by a LLM as constrained reasoner to generate detailed explanations for detected hallucinated content. This study optimizes the real-time interpretable hallucination detection by introducing effective prompting techniques that align LLM-generated explanations with SLM decisions. Empirical experiment results demonstrate its effectiveness, thereby enhancing the overall user experience.
Recent studies show that large language models (LLMs) struggle with technical standards in telecommunications. We propose a fine-tuned retrieval-augmented generation (RAG) system based on the Phi-2 small language model (SLM) to serve as an oracle for communication networks. Our developed system leverages forward-looking semantic chunking to adaptively determine parsing breakpoints based on embedding similarity, enabling effective processing of diverse document formats. To handle the challenge of multiple similar contexts in technical standards, we employ a re-ranking algorithm to prioritize the most relevant retrieved chunks. Recognizing the limitations of Phi-2's small context window, we implement a recent technique, namely SelfExtend, to expand the context window during inference, which not only boosts the performance but also can accommodate a wider range of user queries and design requirements from customers to specialized technicians. For fine-tuning, we utilize the low-rank adaptation (LoRA) technique to enhance computational efficiency during training and enable effective fine-tuning on small datasets. Our comprehensive experiments demonstrate substantial improvements over existing question-answering approaches in the telecom domain, achieving performance that exceeds larger language models such as GPT-4 (which is about 880 times larger in size). This work presents a novel approach to leveraging SLMs for communication networks, offering a balance of efficiency and performance. This work can serve as a foundation towards agentic language models for networks.
With the rapid development of Text-to-Image (T2I) models, biases in human image generation against demographic social groups become a significant concern, impacting fairness and ethical standards in AI. Some researchers propose their methods to tackle with the issue. However, existing methods are designed for specific models with fixed prompts, limiting their adaptability to the fast-evolving models and diverse practical scenarios. Moreover, they neglect the impact of hallucinations, leading to discrepancies between expected and actual results. To address these issues, we introduce VersusDebias, a novel and universal debiasing framework for biases in arbitrary T2I models, consisting of an array generation (AG) module and an image generation (IG) module. The self-adaptive AG module generates specialized attribute arrays to post-process hallucinations and debias multiple attributes simultaneously. The IG module employs a small language model to modify prompts according to the arrays and drives the T2I model to generate debiased images, enabling zero-shot debiasing. Extensive experiments demonstrate VersusDebias's capability to debias any models across gender, race, and age simultaneously. In both zero-shot and few-shot scenarios, VersusDebias outperforms existing methods, showcasing its exceptional utility. Our work is accessible at https://github.com/VersusDebias/VersusDebias to ensure reproducibility and facilitate further research.
Recent advancements in large language models have heavily relied on the large reward model from reinforcement learning from human feedback for fine-tuning. However, the use of a single reward model across various domains may not always be optimal, often requiring retraining from scratch when new domain data is introduced. To address these challenges, we explore the utilization of small language models operating in a domain-specific manner based on router mechanisms. Our three approaches are: 1) utilize mixture of experts to form a single reward model by modularizing an internal router and experts, 2) employing external router to select the appropriate reward model from multiple domain-specific models, and 3) the framework reduces parameter size by loading reward models and router adapters onto a single small language model using adapters. Experimental validation underscores the effectiveness of our approach, demonstrating performance comparable to baseline methods while also reducing the total parameter size.
Speculative decoding has emerged as a promising technique to accelerate the inference of Large Language Models (LLMs) by employing a small language model to draft a hypothesis sequence, which is then validated by the LLM. The effectiveness of this approach heavily relies on the balance between performance and efficiency of the draft model. In our research, we focus on enhancing the proportion of draft tokens that are accepted to the final output by generating multiple hypotheses instead of just one. This allows the LLM more options to choose from and select the longest sequence that meets its standards. Our analysis reveals that hypotheses produced by the draft model share many common token sequences, suggesting a potential for optimizing computation. Leveraging this observation, we introduce an innovative approach utilizing a directed acyclic graph (DAG) to manage the drafted hypotheses. This structure enables us to efficiently predict and merge recurring token sequences, vastly reducing the computational demands of the draft model. We term this approach Graph-structured Speculative Decoding (GSD). We apply GSD across a range of LLMs, including a 70-billion parameter LLaMA-2 model, and observe a remarkable speedup of 1.73$\times$ to 1.96$\times$, significantly surpassing standard speculative decoding.
Recently, various encoder-only and encoder-decoder pre-trained models like BERT and T5 have been applied to automatic essay scoring (AES) as small language models. However, existing studies have primarily treated this task akin to a classification problem, focusing solely on outputting scores in the target text without offering interpretations for the generated scores. Departing from the approaches, we introduce Reasoning Distillation-Based Evaluation (RDBE), which integrates interpretability to elucidate the rationale behind model scores while enhancing performance through initial reasoning. This interpretive capability is acquired during training by leveraging generated reasoning from a large language model (LLM) to distill a small language model (SLM). Our experimental results demonstrate the efficacy of RDBE across all scoring rubrics considered in the dataset. RDBE outperforms both zero-shot LLM generation and generation from a baseline fine-tuned model, establishing itself as state-of-the-art in the corresponding dataset. This highlights its practical interpretative output and enhanced performance.
The goal of text style transfer is to transform the style of texts while preserving their original meaning, often with only a few examples of the target style. Existing style transfer methods generally rely on the few-shot capabilities of large language models or on complex controllable text generation approaches that are inefficient and underperform on fluency metrics. We introduce TinyStyler, a lightweight but effective approach, which leverages a small language model (800M params) and pre-trained authorship embeddings to perform efficient, few-shot text style transfer. We evaluate on the challenging task of authorship style transfer and find TinyStyler outperforms strong approaches such as GPT-4. We also evaluate TinyStyler's ability to perform text attribute style transfer (formal $\leftrightarrow$ informal) with automatic and human evaluations and find that the approach outperforms recent controllable text generation methods. Our model has been made publicly available at https://huggingface.co/tinystyler/tinystyler .
In the context of real-world applications, leveraging large language models (LLMs) for domain-specific tasks often faces two major challenges: domain-specific knowledge privacy and constrained resources. To address these issues, we propose PDSS, a privacy-preserving framework for step-by-step distillation of LLMs. PDSS works on a server-client architecture, wherein client transmits perturbed prompts to the server's LLM for rationale generation. The generated rationales are then decoded by the client and used to enrich the training of task-specific small language model(SLM) within a multi-task learning paradigm. PDSS introduces two privacy protection strategies: the Exponential Mechanism Strategy and the Encoder-Decoder Strategy, balancing prompt privacy and rationale usability. Experiments demonstrate the effectiveness of PDSS in various text generation tasks, enabling the training of task-specific SLM with enhanced performance while prioritizing data privacy protection.