Search

Natural Language Inference (NLI) remains an important benchmark task for LLMs. NLI datasets are a springboard for transfer learning to other semantic tasks, and NLI models are standard tools for identifying the faithfulness of model-generated text. There are several large scale NLI datasets today, and models have improved greatly by hill-climbing on these collections. Yet their realistic performance on out-of-distribution/domain data is less well-understood. We explore the opportunity for synthetic high-quality datasets to adapt NLI models for zero-shot use in downstream applications across new and unseen text domains. We demonstrate a new approach for generating NLI data in diverse domains and lengths, so far not covered by existing training sets. The resulting examples have meaningful premises, the hypotheses are formed in creative ways rather than simple edits to a few premise tokens, and the labels have high accuracy. We show that models trained on this data ($685$K synthetic examples) have the best generalization to completely new downstream test settings. On the TRUE benchmark, a T5-small model trained with our data improves around $7\%$ on average compared to training on the best alternative dataset. The improvements are more pronounced for smaller models, while still meaningful on a T5 XXL model. We also demonstrate gains on test sets when in-domain training data is augmented with our domain-general synthetic data.

Despite the potential impact of explanations on decision making, there is a lack of research on quantifying their effect on users' choices. This paper presents an experimental protocol for measuring the degree to which positively or negatively biased explanations can lead to users choosing suboptimal recommendations. Key elements of this protocol include a preference elicitation stage to allow for personalizing recommendations, manual identification and extraction of item aspects from reviews, and a controlled method for introducing bias through the combination of both positive and negative aspects. We study explanations in two different textual formats: as a list of item aspects and as fluent natural language text. Through a user study with 129 participants, we demonstrate that explanations can significantly affect users' selections and that these findings generalize across explanation formats., Comment: Extended Abstracts of the 2023 CHI Conference on Human Factors in Computing Systems (CHI EA '23), 2023

The classical machine learning paradigm requires the aggregation of user data in a central location where machine learning practitioners can preprocess data, calculate features, tune models and evaluate performance. The advantage of this approach includes leveraging high performance hardware (such as GPUs) and the ability of machine learning practitioners to do in depth data analysis to improve model performance. However, these advantages may come at a cost to data privacy. User data is collected, aggregated, and stored on centralized servers for model development. Centralization of data poses risks, including a heightened risk of internal and external security incidents as well as accidental data misuse. Federated learning with differential privacy is designed to avoid the server-side centralization pitfall by bringing the ML learning step to users' devices. Learning is done in a federated manner where each mobile device runs a training loop on a local copy of a model. Updates from on-device models are sent to the server via encrypted communication and through differential privacy to improve the global model. In this paradigm, users' personal data remains on their devices. Surprisingly, model training in this manner comes at a fairly minimal degradation in model performance. However, federated learning comes with many other challenges due to its distributed nature, heterogeneous compute environments and lack of data visibility. This paper explores those challenges and outlines an architectural design solution we are exploring and testing to productionize federated learning at Meta scale.

Small cations such as guanidinium and cesium can act as templating cations to form low dimensional phases (2D, 1D, 0D) in the case of excess of organic halides. However, such phases with the widely used formamidinium (FA+) cation have not been reported so far. In this study, we discovered two novel low dimensional phases with a composition of FA2PbBr4 and investigated the prerequisites of their formation upon crystallization of FABr-excessive solutions of FAPbBr3. We found that both phases have the structure of (110) layered perovskite but are represented by two different polymorphs with eclipsed and staggered arrangement of adjacent layers. It was shown that FA2PbBr4 phases usually exist in a labile equilibrium with FAPbBr3 3D perovskite and can form composites with it. The optical properties of both polymorphs were comprehensively studied by means of absorption spectroscopy, diffuse reflection spectroscopy and photoluminescence spectroscopy. DFT calculations were applied to investigate the band structure of the FA2PbBr4 and to corroborate the conclusions on their optoelectronic properties. As a result, we found that FA2PbBr4 phases irradiated by UV can exhibit effective green photoluminescence due to a transfer of excitation energy to defective states or 3D perovskite inclusions., Comment: 24 pages, 19 figures

We describe a first open-access database of experimentally investigated hybrid organic-inorganic materials with two-dimensional (2D) perovskite-like crystal structure. The database includes 515 compounds, containing 180 different organic cations, 10 metals (Pb, Sn, Bi, Cd, Cu, Fe, Ge, Mn, Pd, Sb) and 3 halogens (I, Br, Cl) known so far and will be regularly updated. The database contains a geometrical and crystal chemical analysis of the structures, which are useful to reveal quantitative structure-property relationships for this class of compounds. We show that the penetration depth of spacer organic cation into the inorganic layer and M-X-M bond angles increase in the number of inorganic layers (n). The machine learning model is developed and trained on the database, for the prediction of a band gap with accuracy within 0.1 eV. Another machine learning model is trained for the prediction of atomic partial charges with accuracy within 0.01 e. We show that the predicted values of band gaps decrease with an increase of the n and with an increase of M-X-M angles for single-layered perovskites. In general, the proposed database and machine learning models are shown to be useful tools for the rational design of new 2D hybrid perovskite materials.