Recently, many language models have been pretrained with a "midtraining" phase, in which higher quality, often instruction-formatted data, is mixed in at the end of pretraining. Despite the popularity of this practice, there is little scientific understanding of this phase of model training or why it is effective. In this work, we conduct the first systematic investigation of midtraining through controlled experiments with language models pretrained from scratch and fine-tuned on supervised finetuning datasets in different domains. We find that when compared after supervised fine-tuning, the effectiveness of midtraining is highest in the math and code domains, where midtraining can best reduce the syntactic gap between pretraining and posttraining data. In these cases, midtraining consistently outperforms continued pretraining in both in-domain validation loss as well as pretraining data forgetting after posttraining. We conduct ablations on the starting time of the midtraining phase and mixture weights of the midtraining data, using code midtraining as a case study, and find that timing has a greater impact than mixture weights, with earlier introduction of specialized data, yielding greater benefits in-domain as well as preserving general language modeling better. These findings establish midtraining as a domain adaptation technique that compared to continued pretraining yields better performance through reduced forgetting.

Data-driven decision making relies on estimating the effect of interventions, i.e. causal effect estimation . For example, a doctor must decide which medicine she will give her patient, ideally the one with the greatest effect on positive outcomes.

Hostile attribution bias is the tendency to interpret social interactions as intentionally hostile. The Ambiguous Intentions Hostility Questionnaire (AIHQ) is commonly used to measure hostile attribution bias, and includes open-ended questions where participants describe the perceived intentions behind a negative social situation and how they would respond. While these questions provide insights into the contents of hostile attributions, they require time-intensive scoring by human raters. In this study, we assessed whether large language models can automate the scoring of AIHQ open-ended responses. We used a previously collected dataset in which individuals with traumatic brain injury (TBI) and healthy controls (HC) completed the AIHQ and had their open-ended responses rated by trained human raters. We used half of these responses to fine-tune the two models on human-generated ratings, and tested the fine-tuned models on the remaining half of AIHQ responses. Results showed that model-generated ratings aligned with human ratings for both attributions of hostility and aggression responses, with fine-tuned models showing higher alignment. This alignment was consistent across ambiguous, intentional, and accidental scenario types, and replicated previous findings on group differences in attributions of hostility and aggression responses between TBI and HC groups. The fine-tuned models also generalized well to an independent nonclinical dataset. To support broader adoption, we provide an accessible scoring interface that includes both local and cloud-based options. Together, our findings suggest that large language models can streamline AIHQ scoring in both research and clinical contexts, revealing their potential to facilitate psychological assessments across different populations.

The field of text privatization often leverages the notion of $\textit{Differential Privacy}$ (DP) to provide formal guarantees in the rewriting or obfuscation of sensitive textual data. A common and nearly ubiquitous form of DP application necessitates the addition of calibrated noise to vector representations of text, either at the data- or model-level, which is governed by the privacy parameter $\varepsilon$. However, noise addition almost undoubtedly leads to considerable utility loss, thereby highlighting one major drawback of DP in NLP. In this work, we introduce a new sentence infilling privatization technique, and we use this method to explore the effect of noise in DP text rewriting. We empirically demonstrate that non-DP privatization techniques excel in utility preservation and can find an acceptable empirical privacy-utility trade-off, yet cannot outperform DP methods in empirical privacy protections. Our results highlight the significant impact of noise in current DP rewriting mechanisms, leading to a discussion of the merits and challenges of DP in NLP, as well as the opportunities that non-DP methods present.

Evaluating large language models (LLMs) is costly: it requires the generation and examination of LLM outputs on a large-scale benchmark of various tasks. This paper investigates how to efficiently reduce the tasks used to benchmark LLMs without affecting the evaluation quality. Our study reveals that task transferability and relevance provide critical information to identify the most representative subset of tasks via optimizing a facility location function. We propose a practically efficient metric for estimating the transferability between two tasks via in-context learning (ICL). By analyzing the pairwise transferability, we can reduce tasks in a modern LLM benchmark (e.g., MMLU or FLAN) to 5% while inducing only a <4% difference to the evaluation on the original benchmark. Compared to prior works, our method is training-free, gradient-free, and highly efficient requiring ICL only.

There is a need for empathetic and coherent responses in automated chatbot-facilitated psychotherapy sessions. This study addresses the challenge of enhancing the emotional and contextual understanding of large language models (LLMs) in psychiatric applications. We introduce a novel framework that integrates multiple emotion lexicons, including NRC Emotion Lexicon, VADER, WordNet, and SentiWordNet, with state-of-the-art LLMs such as LLAMA 2, Flan-T5, ChatGPT 3.0, and ChatGPT 4.0. The primary dataset comprises over 2,000 therapy session transcripts from the Counseling and Psychotherapy database, covering discussions on anxiety, depression, trauma, and addiction. We segment the transcripts into smaller chunks, enhancing them with lexical features and computing embeddings using BERT, GPT-3, and RoBERTa to capture semantic and emotional nuances. These embeddings are stored in a FAISS vector database, enabling efficient similarity search and clustering based on cosine similarity. Upon user query, the most relevant segments are retrieved and provided as context to the LLMs, significantly improving the models' ability to generate empathetic and contextually appropriate responses. Experimental evaluations demonstrate that in-corporating emotion lexicons enhances empathy, coherence, informativeness, and fluency scores. Our findings highlight the critical role of emotional embeddings in improving LLM performance for psychotherapy.

Large language models are able to learn new tasks in context, where they are provided with instructions and a few annotated examples. However, the effectiveness of in-context learning is dependent on the provided context, and the performance on a downstream task can vary considerably, depending on the instruction. Importantly, such dependency on the context can surface in unpredictable ways, e.g., a seemingly more informative instruction might lead to a worse performance. In this paper, we propose an alternative approach, which we term In-Context Probing (ICP). Similar to in-context learning, we contextualize the representation of the input with an instruction, but instead of decoding the output prediction, we probe the contextualized representation to predict the label. Through a series of experiments on a diverse set of classification tasks, we show that in-context probing is significantly more robust to changes in instructions. We further show that ICP performs competitive or superior to finetuning and can be particularly helpful to build classifiers on top of smaller models, with less than a hundred training examples.

Using in-context learning (ICL) for data generation, techniques such as Self-Instruct (Wang et al., 2023) or the follow-up Alpaca (Taori et al., 2023) can train strong conversational agents with only a small amount of human supervision. One limitation of these approaches is that they resort to very large language models (around 175B parameters) that are also proprietary and non-public. Here we explore the application of such techniques to language models that are much smaller (around 10B--40B parameters) and have permissive licenses. We find the Self-Instruct approach to be less effective at these sizes and propose new ICL methods that draw on two main ideas: (a) Categorization and simplification of the ICL templates to make prompt learning easier for the LM, and (b) Ensembling over multiple LM outputs to help select high-quality synthetic examples. Our algorithm leverages the 175 Self-Instruct seed tasks and employs separate pipelines for instructions that require an input and instructions that do not. Empirical investigations with different LMs show that: (1) Our proposed method yields higher-quality instruction tuning data than Self-Instruct, (2) It improves performances of both vanilla and instruction-tuned LMs by significant margins, and (3) Smaller instruction-tuned LMs generate more useful outputs than their larger un-tuned counterparts. Our codebase is available at https://github.com/IBM/ensemble-instruct.

Large language models (LLMs) with instruction fine-tuning demonstrate superior generative capabilities. However, these models are resource-intensive. To alleviate this issue, we explore distilling knowledge from instruction-tuned LLMs into much smaller ones. To this end, we carefully develop a large set of 2.58M instructions based on both existing and newly-generated instructions. In addition to being sizable, we design our instructions to cover a broad set of topics to ensure diversity. Extensive analysis of our instruction dataset confirms its diversity, and we generate responses for these instructions using gpt-3.5-turbo. Leveraging these instructions, we fine-tune a diverse herd of models, collectively referred to as LaMini-LM, which includes models from both the encoder-decoder and decoder-only families, with varying sizes. We evaluate the performance of our models using automatic metrics on 15 different natural language processing (NLP) benchmarks, as well as through human assessment. The results demonstrate that our proposed LaMini-LM models are comparable to competitive baselines, while being nearly 10 times smaller in size.