query/llm_post/X-ROBUST-009/harness/interface.md

X-ROBUST-009 interface

Entry point: model.py

def load(checkpoint_dir: str, models_dir: str, device: str):
    """Return whatever predict() needs; roberta-base is under models_dir/roberta-base."""

def predict(state, texts: list[str]):
    """Return (labels, confidence, abstain), each of length len(texts):
    labels      int 0..149, your best in-scope intent (always given, also when you abstain)
    confidence  float in [0, 1], your probability that `labels` is the right answer
    abstain     bool, True if the assistant should not answer"""

• Intent indices follow $ABENCH_DATA_DIR/intents.json; label −1 in the public data marks out-of-scope queries.
• train.py --data "$ABENCH_DATA_DIR" --out checkpoint must reproduce the checkpoint.

Evaluator code you can read

stages.py (batching, timing), metrics.py (AURC, ECE, AUROC, decision accuracy), textpack.py, abench_metrics.py, abench_stage.py, selfcheck.py, and the reference submission's code under baseline/.

Available software

Python 3, PyTorch (CUDA), transformers, NumPy, SciPy, scikit-learn. Nothing can be installed.