Children from bilingual backgrounds benefit from interactions with parents and teachers to re-acquire their heritage language. In this paper, we investigate how this insight from behavioral study can be incorporated into the learning of small-scale language models. We introduce BAMBINO-LM, a continual pre-training strategy for BabyLM that uses a novel combination of alternation and PPO-based perplexity reward induced from a parent Italian model. Upon evaluation on zero-shot classification tasks for English and Italian, BAMBINO-LM improves the Italian language capability of a BabyLM baseline. Our ablation analysis demonstrates that employing both the alternation strategy and PPO-based modeling is key to this effectiveness gain. We also show that, as a side effect, the proposed method leads to a similar degradation in L1 effectiveness as human children would have had in an equivalent learning scenario. Through its modeling and findings, BAMBINO-LM makes a focused contribution to the pre-training of small-scale language models by first developing a human-inspired strategy for pre-training and then showing that it results in behaviours similar to that of humans.

Large Language Models (LLMs) have demonstrated remarkable capabilities in solving various tasks, yet they often struggle with comprehensively addressing complex and vague problems. Existing approaches, including multi-agent LLM systems, offer solutions to certain challenges but still require manual setup and lack scalability. To address this gap, we propose a novel approach leveraging decomposition to enable LLMs to tackle vague problems effectively. Our approach involves an orchestrating LLM that interacts with users to understand the problem and then decomposes it into tangible sub-problems. Instead of expecting the LLM to solve the entire problem in one go, we train it to ask follow-up questions to gain a deeper understanding of the user's requirements. Once the problem is adequately understood, the orchestrating LLM divides it into smaller, manageable sub-problems. Each sub-problem is then assigned to specialized LLM agents or non-LLM functions for resolution. These agents work in parallel to solve their respective sub-problems, with the orchestrating LLM overseeing the process and compiling the solutions into a comprehensive answer for the user. By adopting this decomposition approach, we alleviate the constraints imposed by token limitations on LLM outputs and empower them to provide nuanced solutions to complex and ambiguous problems. Through our approach, we aim to enable LLMs to think and operate more like humans, breaking down complex problems into manageable parts and collaboratively solving them. This not only enhances the problem-solving capabilities of LLMs but also offers a scalable and efficient method for addressing a wide range of real-world challenges.

One of the most critical tasks of Microsoft sellers is to meticulously track and nurture potential business opportunities through proactive engagement and tailored solutions. Recommender systems play a central role to help sellers achieve their goals. In this paper, we present a content recommendation model which surfaces various types of content (technical documentation, comparison with competitor products, customer success stories etc.) that sellers can share with their customers or use for their own self-learning. The model operates at the opportunity level which is the lowest possible granularity and the most relevant one for sellers. It is based on semantic matching between metadata from the contents and carefully selected attributes of the opportunities. Considering the volume of seller-managed opportunities in organizations such as Microsoft, we show how to perform efficient semantic matching over a very large number of opportunity-content combinations. The main challenge is to ensure that the top-5 relevant contents for each opportunity are recommended out of a total of $\approx 40,000$ published contents. We achieve this target through an extensive comparison of different model architectures and feature selection. Finally, we further examine the quality of the recommendations in a quantitative manner using a combination of human domain experts as well as by using the recently proposed "LLM as a judge" framework.

The pervasive deployment of Large Language Models-LLMs in various sectors often neglects the nuanced requirements of individuals and small organizations, who benefit more from models precisely tailored to their specific business contexts rather than those with broadly superior general capabilities. This work introduces \textbf{AnyTaskTune}, a novel fine-tuning methodology coined as \textbf{Task-Fine-Tune}, specifically developed to elevate model performance on a diverse array of domain-specific tasks. This method involves a meticulous process to identify and define targeted sub-tasks within a domain, followed by the creation of specialized enhancement datasets for fine-tuning, thereby optimizing task-specific model performance. We conducted comprehensive fine-tuning experiments not only in the legal domain for tasks such as keyword extraction and sentence prediction but across over twenty different sub-tasks derived from the domains of finance, healthcare, law, psychology, consumer services, and human resources. To substantiate our approach and facilitate community engagement, we will open-source these bilingual task datasets. Our findings demonstrate that models fine-tuned using the \textbf{Task-Fine-Tune} methodology not only achieve superior performance on these specific tasks but also significantly outperform models with higher general capabilities in their respective domains. Our work is publicly available at \url{https://github.com/PandaVT/DataTager}.

Among the widely used parameter-efficient fine-tuning (PEFT) methods, LoRA and its variants have gained considerable popularity because of avoiding additional inference costs. However, there still often exists an accuracy gap between these methods and full fine-tuning (FT). In this work, we first introduce a novel weight decomposition analysis to investigate the inherent differences between FT and LoRA. Aiming to resemble the learning capacity of FT from the findings, we propose Weight-Decomposed Low-Rank Adaptation (DoRA). DoRA decomposes the pre-trained weight into two components, magnitude and direction, for fine-tuning, specifically employing LoRA for directional updates to efficiently minimize the number of trainable parameters. By employing \ours, we enhance both the learning capacity and training stability of LoRA while avoiding any additional inference overhead. \ours~consistently outperforms LoRA on fine-tuning LLaMA, LLaVA, and VL-BART on various downstream tasks, such as commonsense reasoning, visual instruction tuning, and image/video-text understanding. Code is available at https://github.com/NVlabs/DoRA.

Large language models (LLMs) present an enormous evolution in the strategic potential of conversational recommender systems (CRS). Yet to date, research has predominantly focused upon technical frameworks to implement LLM-driven CRS, rather than end-user evaluations or strategic implications for firms, particularly from the perspective of a small to medium enterprises (SME) that makeup the bedrock of the global economy. In the current paper, we detail the design of an LLM-driven CRS in an SME setting, and its subsequent performance in the field using both objective system metrics and subjective user evaluations. While doing so, we additionally outline a short-form revised ResQue model for evaluating LLM-driven CRS, enabling replicability in a rapidly evolving field. Our results reveal good system performance from a user experience perspective (85.5% recommendation accuracy) but underscore latency, cost, and quality issues challenging business viability. Notably, with a median cost of $0.04 per interaction and a latency of 5.7s, cost-effectiveness and response time emerge as crucial areas for achieving a more user-friendly and economically viable LLM-driven CRS for SME settings. One major driver of these costs is the use of an advanced LLM as a ranker within the retrieval-augmented generation (RAG) technique. Our results additionally indicate that relying solely on approaches such as Prompt-based learning with ChatGPT as the underlying LLM makes it challenging to achieve satisfying quality in a production environment. Strategic considerations for SMEs deploying an LLM-driven CRS are outlined, particularly considering trade-offs in the current technical landscape.

When asked to summarize articles or answer questions given a passage, large language models (LLMs) can hallucinate details and respond with unsubstantiated answers that are inaccurate with respect to the input context. This paper describes a simple approach for detecting such contextual hallucinations. We hypothesize that contextual hallucinations are related to the extent to which an LLM attends to information in the provided context versus its own generations. Based on this intuition, we propose a simple hallucination detection model whose input features are given by the ratio of attention weights on the context versus newly generated tokens (for each attention head). We find that a linear classifier based on these lookback ratio features is as effective as a richer detector that utilizes the entire hidden states of an LLM or a text-based entailment model. The lookback ratio-based detector -- Lookback Lens -- is found to transfer across tasks and even models, allowing a detector that is trained on a 7B model to be applied (without retraining) to a larger 13B model. We further apply this detector to mitigate contextual hallucinations, and find that a simple classifier-guided decoding approach is able to reduce the amount of hallucination, for example by 9.6% in the XSum summarization task.

Large language models (LLMs) have been shown to propagate and amplify harmful stereotypes, particularly those that disproportionately affect marginalised communities. To understand the effect of these stereotypes more comprehensively, we introduce GlobalBias, a dataset of 876k sentences incorporating 40 distinct gender-by-ethnicity groups alongside descriptors typically used in bias literature, which enables us to study a broad set of stereotypes from around the world. We use GlobalBias to directly probe a suite of LMs via perplexity, which we use as a proxy to determine how certain stereotypes are represented in the model's internal representations. Following this, we generate character profiles based on given names and evaluate the prevalence of stereotypes in model outputs. We find that the demographic groups associated with various stereotypes remain consistent across model likelihoods and model outputs. Furthermore, larger models consistently display higher levels of stereotypical outputs, even when explicitly instructed not to.

Training large language models (LLMs) in low-resource languages such as Hebrew poses unique challenges. In this paper, we introduce DictaLM2.0 and DictaLM2.0-Instruct, two LLMs derived from the Mistral model, trained on a substantial corpus of approximately 200 billion tokens in both Hebrew and English. Adapting a pre-trained model to a new language involves specialized techniques that differ significantly from training a model from scratch or further training existing models on well-resourced languages such as English. We outline these novel training methodologies, which facilitate effective learning and adaptation to the linguistic properties of Hebrew. Additionally, we fine-tuned DictaLM2.0-Instruct on a comprehensive instruct dataset to enhance its performance on task-specific instructions. To rigorously evaluate our models, we introduce a new benchmark suite for Hebrew LLM evaluation, covering a diverse set of tasks including Question Answering, Sentiment Analysis, Winograd Schema Challenge, Translation, and Summarization. Our work not only addresses the intricacies of training LLMs in low-resource languages but also proposes a framework that can be leveraged for adapting other LLMs to various non-English languages, contributing to the broader field of multilingual NLP.

In our experiments, we start on top of a 15B parameter LM that has seen 8T tokens of pretraining Language modeling abilities have seen massive data (Parmar et al., 2024). Experimenting with a improvements over the past few years (Brown well trained model of this scale ensures that our et al., 2020; Chowdhery et al., 2022; OpenAI, 2024; findings will be transferable to most settings and Team, 2024). While these advancements have enabled model sizes. We first identify the type of data distribution language models (LMs) to become highlyskilled that should be used during continued pretraining conversational agents (OpenAI, 2024; Anthropic, and find that it is optimal to have two distributions, 2024; Team, 2024), they have come with with the final one more heavily weighting increased computational cost as pretraining has become data sources that relate to the abilities we want to ever more expensive due to both the number improve in the model. Second, we determine what of model parameters (Team et al., 2024; DeepSeek-learning rate schedules enable the most efficient AI et al., 2024) and pretraining dataset size (Touvron learning during continued pretraining and determine et al., 2023; Gemma Team, 2024; Parmar et al., that the most performant one strikes a balance 2024) continuing to grow in scale. With new LMs between magnitude of learning rate and steepness that set state of the art accuracy being released of decay. Lastly, we show how the learning rate on a frequent basis, LMs developed only a couple value at which we switch between data distributions months back are becoming obsolete as their affects downstream accuracy and identify the capabilities are no longer up to par. This leaves point at which this switch should be made.