arpit-bhayani/rush

Checkpoints

This directory stores calibrated risk parameters and trained neural head checkpoints for Rush.

Tracked Files

• conformal_calibration.json: Calibrated non-conformity quantiles and risk thresholds evaluated on held-out test splits (e.g. Banking77).

Binary Weights (*.pt - Ignored from Git)

The following binary weights are kept locally or downloaded from Hugging Face Hub / GitHub Releases:

• semantic_option_bank.pt: Trained SemanticOptionBank metric tensor ($W$) and temperature ($\tau$) on Banking77.
• span_pointer.pt: Trained SpanPointerHead start/end boundary projection layers on SQuAD.

Reproducing Checkpoints Locally

To train or re-calibrate the checkpoints from scratch:

# 1. Train Semantic Option Bank on Banking77 and calibrate conformal quantiles
python3 rush/train/train_banking77.py

# 2. Train Extractive Span Pointer Head on SQuAD
python3 rush/train/train_span_pointer.py