mobilebert-uncased-squad-v2

Performs extractive QA by encoding question-passage pairs through a 24-layer MobileBERT transformer architecture, then predicting start and end token positions via dense classification heads. Uses SQuAD v2 fine-tuning which includes unanswerable questions, enabling the model to abstain when no valid answer exists in the passage. The model outputs logit scores for each token position, with post-processing to extract the highest-confidence span.

Leverages SQuAD v2 training which includes ~33% unanswerable questions to learn when to abstain from answering. The model predicts a special [CLS] token logit score alongside span predictions; when this score exceeds the span confidence, the model returns 'unanswerable' rather than forcing an incorrect extraction. This is implemented as a three-way classification: start position, end position, and 'no answer' token probability.

Model is distributed in multiple optimized formats: PyTorch (.pt), ONNX (.onnx for cross-platform inference), and SafeTensors (.safetensors for secure deserialization). ONNX format enables hardware-accelerated inference on mobile (iOS/Android via ONNX Runtime), browsers (WebAssembly), and edge devices. The 25MB base model can be further quantized (INT8, FP16) reducing size to 6-12MB with <5% accuracy loss, enabling deployment on devices with <100MB storage.

Supports batched inference through HuggingFace transformers pipeline API, which handles tokenization, padding, and attention mask generation automatically. Uses dynamic padding (pads to max length in batch, not fixed 512) to reduce computation. Attention mechanism is standard multi-head self-attention (12 heads in MobileBERT) with token-level masking to ignore padding tokens, enabling efficient processing of variable-length questions and passages.

MobileBERT was trained using knowledge distillation from BERT-large as the teacher model, transferring learned representations into a smaller student architecture. This enables fine-tuning on downstream tasks (like SQuAD v2) with minimal accuracy loss despite 4.3x parameter reduction. The distillation approach uses intermediate layer matching and attention transfer, not just final logit matching, preserving semantic understanding across layers.

Model is distributed in three formats: PyTorch (.pt), ONNX (.onnx), and SafeTensors (.safetensors). SafeTensors is a newer format that avoids pickle deserialization vulnerabilities by using a simple binary format with explicit type information. This enables safe loading of untrusted model files without arbitrary code execution risk. All three formats are available from the HuggingFace Hub with automatic format detection.

Model is compatible with Azure ML inference endpoints, enabling serverless QA deployment with automatic scaling. Azure integration includes model registration, endpoint creation, and REST API exposure without manual infrastructure setup. The model can be deployed as a managed endpoint with auto-scaling based on request volume, with built-in monitoring and logging.