vi-mrc-large

Performs extractive QA by fine-tuned RoBERTa-large encoder that predicts start and end token positions within a passage to extract answer spans. Uses transformer-based sequence classification with token-level logits to identify answer boundaries, trained on Vietnamese SQuAD-format datasets with cross-lingual transfer from English pre-training. Architecture leverages masked language modeling representations to contextualize Vietnamese text and identify semantically relevant answer spans without generating new text.

Leverages RoBERTa-large's multilingual pre-training (trained on 100+ languages including Vietnamese and English) to transfer knowledge from English SQuAD fine-tuning to Vietnamese QA tasks. The model architecture preserves language-agnostic contextual representations learned during pre-training, allowing the token classification head to generalize across Vietnamese and English without explicit cross-lingual alignment. Fine-tuning on Vietnamese SQuAD data adapts the shared encoder representations while maintaining transfer benefits from English QA patterns.

Supports standard SQuAD format input/output (JSON with passages, questions, answers with character offsets) for both training and evaluation. The model integrates with HuggingFace Datasets library to load SQuAD-compatible data, compute exact-match and F1 metrics during training, and enable reproducible benchmarking. Fine-tuning pipeline handles tokenization, token-to-character offset mapping, and loss computation for span prediction without requiring custom data loaders.

Outputs logit scores for start and end token positions, enabling confidence-based answer filtering and ranking. The model computes softmax probabilities over all tokens in the passage for both start and end positions, allowing downstream systems to rank candidate answers by joint probability (start_prob × end_prob) or filter low-confidence predictions. This enables uncertainty quantification and selective answer suppression in production systems.

Supports efficient batch processing of multiple passage-question pairs through HuggingFace Transformers pipeline API, which handles tokenization, batching, and output aggregation. The model processes variable-length passages and questions by padding to max sequence length within each batch, enabling GPU-accelerated inference across multiple examples. Batch size can be tuned for memory/latency tradeoffs on different hardware.

Model is compatible with Azure ML endpoints for serverless inference deployment, enabling pay-per-use QA without managing infrastructure. Azure integration handles model versioning, auto-scaling based on request volume, and REST API exposure. The model can be deployed as a managed endpoint with configurable compute resources (CPU/GPU), enabling cost-optimized inference for variable traffic patterns.