Learning Structure-Aware Representations of Dependent Types

Authors: Konstantinos Kogkalidis, Orestis Melkonian, Jean-Philippe Bernardy

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We introduce and release a novel dataset of Agda program-proofs...we propose a novel neural architecture targeted at faithfully representing dependently-typed programs...We instantiate and evaluate our architecture in a premise selection setup, where it achieves promising initial results, surpassing strong baselines. and 5 Experiments & Results, Table 1: Model performance under different ablations across all evaluation sets.
Researcher Affiliation Collaboration 1Aalto University 2University of Bologna 3Input Output (IOG) 4University of Gothenburg 5Chalmers University of Technology
Pseudocode No The paper describes its architecture and various components (e.g., Efficient Attention, Structured Attention, Variable Representations) but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes Source code, configuration files and scripts supporting experiment execution are available at https://github.com/konstantinos Kokos/quill.
Open Datasets Yes Concisely, we train using Info NCE on a size-constrained subset of the Agda standard library. The extraction algorithm is developed against version 2.6.3 of Agda5 and applied on version 1.7.2 of the Agda standard library6 an open-source, community-developed collection of definitions and proofs for general purpose coding and proving with Agda. Footnote 6 provides a URL: https://github.com/agda/agda-stdlib/releases/tag/v1.7.2. It also mentions Unimath [Rijke et al.] and Type Topology [Escard o and contributors] Agda libraries for additional evaluation sets, both with citations.
Dataset Splits Yes We fix an 85%-15% training-evaluation split over the 818 files extracted from the Agda standard library. and this amounts to 481 training and 92 evaluation files, counting 15,244 and 4,120 holes respectively.
Hardware Specification Yes We train on a single A100 for 100 epochs
Software Dependencies No The extraction algorithm is developed against version 2.6.3 of Agda5. While Agda's version is specified, other software dependencies like PyTorch, used implicitly through mentions of Adam W and other components, are not explicitly listed with their versions.
Experiment Setup Yes We extensively detail our exact experimental setup in Appendix B. Concisely, we train using Info NCE on a size-constrained subset of the Agda standard library. and Table 2: Hyperparameter setup which includes detailed model and optimization hyperparameters.