Antibody CDR Position Control - Trained Model Checkpoints (Mamba + Transformer)

Complete trained model checkpoints for the manuscript:"Precise CDR Position Control in Antibody Sequence Generation Using Conditional Deep Generative Models" This archive contains the full trained PyTorch model weights for both architectures described in the manuscript: 1. Mamba Model (mamba_epoch8_best_model.pt, 1.0 GB) - Selective State Space Model architecture - 88.6 million parameters - Training: 8 epochs, 52 GPU-hours on 6× A100 40GB - Best validation loss: 0.4636 - CDR position accuracy: 100.0% - Trained on 10.88M antibody sequences from OAS + SAbDab 2. Transformer Model (transformer_epoch13_best_model.pt, 579 MB) - Decoder-only Transformer (GPT-style) architecture - 50.5 million parameters - Training: 13 epochs, 65 GPU-hours on 6× A100 40GB - Best validation loss: 0.6187 - CDR position accuracy: 98.0% - Inference speed: 6.49 seq/s (2.13× faster than Mamba) 3. Configuration Files - vocab.json: 40-token vocabulary mapping - config_mamba.yaml: Complete Mamba architecture specification - config_transformer.yaml: Complete Transformer architecture specification - README.md: Usage instructions and model details Key Results:- Both models achieve statistically significant conditional control over CDR3 properties (p<0.001)- Mamba: Superior training quality (25% lower validation loss)- Transformer: Faster inference (2.13× speedup) These models can be loaded directly in PyTorch for sequence generation and evaluation. All training code, data preprocessing scripts, and usage examples are available in the GitHub repository. Manuscript Status: Under review at PLOS ONE (October 2025)GitHub Repository: https://github.com/261732506/antibody-cdr-position-control Upon publication, this record will be updated with full author information and manuscript DOI.