hiv1-recombinant-classifier

HIV-1 Recombinant Classification

Attention-based Transformer encoder for HIV-1 recombinant/non-recombinant classification.

Model

• Representation: 3-mer tokenization
• Architecture: Transformer Encoder
• Transformer layers: 2
• Attention heads: 8
• Embedding dimension: 128
• Feed-forward dimension: 256
• Maximum input length: 4,395 3-mers
• Classes:
  • Non-Recombinant
  • Recombinant

Deployment Model

The deployment model uses post-training dynamic INT8 quantization of eligible Linear layers.

The model is intended for CPU-oriented inference.

The embedding and Transformer attention components are not fully quantized.

Source

The deployed model originates from:

Experiment 5 — Best Transformer

The quantized model was generated as part of:

Experiment 6 — Post-Training Dynamic INT8 Quantization

Reproducibility

Source FP32 checkpoint SHA256:

2a5641ef378603c6d9f2e98a1a4a29bc428e6774c92ffb80d39f1c90c01bbb83

Dynamic INT8 checkpoint SHA256:

69935da6113edfb1f6eded7ded6bb6b6219b173aa24325b0e26d8da43012b289

Train-only vocabulary SHA256:

1f626e456022bafa8547b925bdd03907fabcb1d3a8fc09353f008b63536ad0ae

Intended Use

This model is a research prototype for HIV-1 genomic surveillance and recombinant classification research.

It is not a clinical diagnostic tool and should not be used for patient diagnosis or clinical decision-making.

Files

• model/Transformer_Encoder_best_DYNAMIC_INT8.pt
  • Deployment model
• model/Transformer_Encoder_best.pt
  • Original FP32 source checkpoint
• tokenizer/TRAIN_ONLY_3MER_VOCABULARY.csv
  • Training-only 3-mer vocabulary
• metadata/model_config.json
  • Architecture and deployment metadata
• metadata/file_manifest.json
  • File integrity manifest

Research Project

HIV-1 Subtype/Recombinant Classification

Attention-Based Deep Learning for HIV-1 Genomic Surveillance and Subtype Classification in Uganda.