Quantization is an indispensable technique for serving Large Language Models (LLMs) and has recently found its way into LoRA fine-tuning (Dettmers et al., 2023). In this work we focus on the scenario where quantization and LoRA fine- tuning are applied together on a pre-trained model. In such cases it is common to observe a consistent gap in the performance on downstream tasks between full fine-tuning and quantization plus LoRA fine-tuning approach. In response, we propose LoftQ (LoRA-Fine-Tuning-aware Quantization), a novel quantization framework that simultaneously quantizes an LLM and finds a proper low-rank initialization for LoRA fine-tuning. Such an initialization alleviates the discrep- ancy between the quantized and full-precision model and significantly improves the generalization in downstream tasks. We evaluate our method on natural lan- guage understanding, question answering, summarization, and natural language generation tasks. Experiments show that our method is highly effective and out- performs existing quantization methods, especially in the challenging 2-bit and 2/4-bit mixed precision regimes. We will release our code.

Rigorously establishing the safety of black-box machine learning models with respect to critical risk measures is important for providing guarantees about the behavior of the model. Recently, a notion of a risk controlling prediction set (RCPS) has been introduced by Bates et. al. (JACM '24) for producing prediction sets that are statistically guaranteed to have low risk from machine learning models. Our method extends this notion to the sequential setting, where we provide guarantees even when the data is collected adaptively, and ensures the risk guarantee is anytime-valid, i.e., simultaneously holds at all time steps. Further, we propose a framework for constructing RCPSes for active labeling, i.e., allowing one to use a labeling policy that chooses whether to query the true label for each received data point, and ensures the expected proportion data points whose labels are queried are below a predetermined label budget. We also describe how to use predictors (e.g., the machine learning model we are providing risk control guarantees for) to further improve the utility of our RCPSes by estimating the expected risk conditioned on the covariates. We characterize the optimal choices of label policy under a fixed label budget, and predictor, and show a regret result that relates the estimation error of the optimal labeling policy and predictor to the wealth process that underlies our RCPSes. Lastly, we present practical ways of formulating label policies and we empirically show that our label policies use fewer labels to reach higher utility than naive baseline labeling strategies on both simulations and real data.

Most language models (LMs) are trained and applied in an autoregressive left-to-right fashion, predicting the next token from the preceding ones. However, this ignores that the full sequence is available during training. In this paper, we introduce ``Meet in the Middle'' (MIM) a new pre-training paradigm that improves data efficiency by training in two directions, left-to-right and right-to-left, and encouraging the respective models to agree on their token distribution for each position. While the primary outcome is an improved left-to-right LM, we also obtain secondary benefits in the infilling task. There, we leverage the two pre-trained directions to propose an infilling procedure that builds the completion simultaneously from both sides. We conduct extensive experiments on both programming and natural languages and show that MIM significantly surpasses existing pre-training paradigms, in both left-to-right generation as well as infilling. Code and models available at https://github.com/microsoft/Meet-in-the-Middle

We develop confidence bounds that hold uniformly over time for off-policy evaluation in the contextual bandit setting. These confidence sequences are based on recent ideas from martingale analysis and are non-asymptotic, non-parametric, and valid at arbitrary stopping times. We provide algorithms for computing these confidence sequences that strike a good balance between computational and statistical efficiency. We empirically demonstrate the tightness of our approach in terms of failure probability and width and apply it to the “gated deployment” problem of safely upgrading a production contextual bandit system.

Virtual support agents have grown in popularity as a way for businesses to provide better and more accessible customer service. Some challenges in this domain include ambiguous user queries as well as changing support topics and user behavior (non-stationarity). We do, however, have access to partial feedback provided by the user (clicks, surveys, and other events) which can be leveraged to improve the user experience. Adaptable learning techniques, like contextual bandits, are a natural fit for this problem setting. In this paper, we discuss real-world implementations of contextual bandits (CB) for the Microsoft virtual agent. It includes intent disambiguation based on neural-linear bandits (NLB) and contextual recommendations based on a collection of multi-armed bandits (MAB). Our solutions have been deployed to production and have improved key business metrics of the Microsoft virtual agent, as confirmed by A/B experiments. Results include a relative increase of over 12% in problem resolution rate and relative decrease of over 4% in escalations to a human operator. While our current use cases focus on intent disambiguation and contextual recommendation for support bots, we believe our methods can be extended to other domains.

We propose an estimator and confidence interval for computing the value of a policy from off-policy data in the contextual bandit setting. To this end we apply empirical likelihood techniques to formulate our estimator and confidence interval as simple convex optimization problems. Using the lower bound of our confidence interval, we then propose an off-policy policy optimization algorithm that searches for policies with large reward lower bound. We empirically find that both our estimator and confidence interval improve over previous proposals in finite sample regimes. Finally, the policy optimization algorithm we propose outperforms a strong baseline system for learning from off-policy data.

The task of Named Entity Recognition (NER) has been well studied under high-resource conditions (e.g., extracting named mentions of PERSON, ORGANIZATION and LOCATION from news articles). However, there are very few studies of the NER task for open-domain collections and in low-resource settings. We focus on NER for low-resource collections, in which any entity types of practical interest to the users of the system must be supported. We try to achieve this with a low cost of annotation of data from the target domain/collection. We propose an entity recognition framework that combines active learning and conditional random fields (CRF), and which provides the flexibility to define new entity types as needed by the users. Our experiments on a help & support corpus show that the system can achieve F1 measure of 0.77 by relying on only 100 manually-annotated sentences.

We create a new online reduction of multiclass classification to binary classification for which training and prediction time scale logarithmically with the number of classes. We show that several simple techniques give rise to an algorithm which is superior to previous logarithmic time classification approaches while competing with one-against-all in space. The core construction is based on using a tree to select a small subset of labels with high recall, which are then scored using a one-against-some structure with high precision.

We propose a novel coding theoretic framework for mitigating stragglers in distributed learning. We show how carefully replicating data blocks and coding across gradients can provide tolerance to failures and stragglers for synchronous Gradient Descent. We implement our schemes in python (using MPI) to run on Amazon EC2, and show how we compare against baseline approaches in running time and generalization error.

This work provides simple algorithms for multi-class (and multi-label) prediction in settings where both the number of examples n and the data dimension d are relatively large. These robust and parameter free algorithms are essentially iterative least-squares updates and very versatile both in theory and in practice. On the theoretical front, we present several variants with convergence guarantees. Owing to their effective use of second-order structure, these algorithms are substantially better than first-order methods in many practical scenarios. On the empirical side, we show how to scale our approach to high dimensional datasets, achieving dramatic computational speedups over popular optimization packages such as Liblinear and Vowpal Wabbit on standard datasets (MNIST and CIFAR-10), while attaining state-of-the-art accuracies.

Representing examples in a way that is compatible with the underlying classifier can greatly enhance the performance of a learning system. In this paper we investigate scalable techniques for inducing discriminative features by taking advantage of simple second order structure in the data. We focus on multiclass classification and show that features extracted from the generalized eigenvectors of the class conditional second moments lead to classifiers with excellent empirical performance. Moreover, these features have attractive theoretical properties, such as inducing representations that are invariant to linear transformations of the input. We evaluate classifiers built from these features on three different tasks, obtaining state of the art results.

Recent advances in high throughput data collection and storage technologies have led to a dramatic increase in the availability of high-resolution time series data sets in various domains. These time series reflect the dynamics of the underlying physical processes in these domains. Detecting changes in a time series over time or changes in the relationships among the time series in a data set containing multiple contemporaneous time series can be useful to detect changes in these physical processes. Contextual events detection algorithms detect changes in the relationships between multiple related time series. In this work, we introduce a new type of contextual events, called group level contextual change events. In contrast to individual contextual change events that reflect the change in behavior of one target time series against a context, group level events reflect the change in behavior of a target group of time series relative to a context group of time series. We propose an online framework to detect two types of group level contextual change events: (i) group formation (i.e., detecting when a set of multiple unrelated timeseries or groups of time series with little prior relationship in their behavior forms a new group of related time series) and (ii) group disbanding (i.e., detecting when one stable set of related time series disbands into two or more subgroups with little relationship in their behavior). We demonstrate this framework using 2 real world datasets and show that the framework detects group level contextual change events that can be explained by plausible causes.

We propose a sampling scheme suitable for reducing a data set prior to selecting a hypothesis with minimum empirical risk. The sampling only considers a subset of the ultimate (unknown) hypothesis set, but can nonetheless guarantee that the final excess risk will compare favorably with utilizing the entire original data set. We demonstrate the practical benefits of our approach on a large dataset which we subsample and subsequently fit with boosted trees.

We cast the problem of identifying basic blocks of code in a binary executable as learning a mapping from a byte sequence to a segmentation of the sequence. In general, inference in segmentation models, such as semi-CRFs, can be cubic in the length of the sequence. By taking advantage of the structure of our problem, we derive a linear-time inference algorithm which makes our approach practical, given that even small programs are tens or hundreds of thousands bytes long. Furthermore, we introduce two loss functions which are appropriate for our problem and show how to use structural SVMs to optimize the learned mapping for these losses. Finally, we present experimental results that demonstrate the advantages of our method against a strong baseline.