PubMed-OCR: PMC Open Access OCR Annotations
A large-scale dataset of 209K+ articles with OCR and layout bounding boxes, enabling layout-aware modeling and document …
Machine-Learning
Molecular Sets (MOSES): A Generative Modeling Benchmark
MOSES provides a standardized benchmarking platform for molecular generative models, featuring datasets, metrics, and …
The Reliability Trap: When 99% Accuracy Isn't Enough
Why high-accuracy LLMs fail in production: exploring the calibration crisis and the challenge of reliable …
The Evolution of Page Stream Segmentation: Rules to LLMs
From brittle rules to reasoning engines: why transformers are the only way to solve the 'Hello World' of document …
ChemBERTa-2: Scaling Molecular Transformers to 77M
Optimizing transformer pretraining for molecules using MLM vs MTR objectives, scaling to 77M compounds from PubChem for …
GP-MoLFormer: Molecular Generation via Transformers
A 46.8M parameter transformer for molecular generation trained on 1.1B SMILES, introducing pair-tuning for efficient …
ChemBERTa: Molecular Property Prediction via Transformers
A systematic evaluation of RoBERTa transformers pretrained on 77M PubChem SMILES for molecular property prediction …
Chemformer: Pre-trained Transformer for Comp Chem
BART-based Transformer pre-trained on 100M molecules using self-supervision to accelerate convergence on chemical …
Building Normalizing Flows with Stochastic Interpolants
A continuous-time normalizing flow using stochastic interpolants and quadratic loss to bypass costly ODE …
Flow Matching for Generative Modeling
A simulation-free framework for training Continuous Normalizing Flows using Conditional Flow Matching and Optimal …
Neural ODEs: Continuous-Depth Deep Learning
Introduces ODE-Nets, a continuous-depth neural network model parameterized by ODEs, enabling constant memory …
Score Matching and Denoising Autoencoders
Theoretical paper proving the equivalence between training Denoising Autoencoders and performing Score Matching on a …