going-doer/Paper2Code
Paper2Code
Paper2Code est une suite d'automatisation de recherche en IA et un pipeline de génération de code par grand modèle de langage, conçus pour transformer des articles de recherche en machine learning en dépôts de code exécutables. Il fonctionne comme un outil pour automatiser la traduction de la littérature scientifique et des descriptions théoriques en implémentations fonctionnelles de machine learning.
Le système utilise un pipeline de génération multi-étapes qui exploite la décomposition document-vers-plan et l'échafaudage automatisé de dépôts pour produire des structures de projet complètes. Il intègre un framework d'évaluation de code automatisé qui utilise une boucle itérative de critique-et-affinement et une évaluation de référence pour noter la correction du code généré par rapport à des dépôts vérifiés.
Le projet couvre plusieurs domaines de capacité, notamment la numérisation d'articles scientifiques, la synthèse automatisée de code et la validation de code de machine learning.
Features
- Program Synthesis Models - Utilizes program synthesis models to translate mathematical descriptions and architectural diagrams into executable code.
- Code Validation Pipelines - Implements code validation pipelines to evaluate the correctness of generated ML implementations via gold-standard comparison.
- Generative Code Models - Leverages generative code models to produce complete repositories based on high-level technical specifications from papers.
- Paper-to-Code Frameworks - Provides a framework for automating the translation of scientific paper descriptions into functional ML implementation and analysis.
- Automated Code Refinement Loops - Provides automated code refinement loops where an LLM critiques and corrects generated code to improve accuracy.
- Paper Digitization Tools - Converts complex theoretical descriptions and equations from academic papers into executable software components.
- Paper-to-Code Implementations - Automates the full pipeline of translating machine learning research papers into functional code implementations.
- Implementation Automation Tools - Functions as an AI research automation suite for planning and generating software directly from scientific literature.
- Code Generation Pipelines - Implements an end-to-end code generation pipeline that transforms research papers into deployable software artifacts.
- Research Generation Pipelines - Employs a multi-stage generation pipeline that sequences planning and analysis phases before triggering code synthesis.
- Feature Decomposition Plans - Decomposes complex research papers into structured feature decomposition plans to guide modular code generation.
- Code Generation Evaluators - Provides a code generation evaluator that scores implementation correctness against reference gold repositories.
- Gold Standard Evaluators - Implements reference-based gold evaluation to score generated code against verified gold repositories.
- Machine Learning Implementations - Generates functional machine learning implementations to reproduce experiments described in scientific literature.
- Project Scaffolders - Implements automated project scaffolders to generate directory hierarchies and dependency files from analyzed research papers.
