FluidInference/verdict-coreml

1"""Tensorize Verdict's native rendered text for one fixed Core ML bucket."""2 3from __future__ import annotations4 5import numpy as np6 7 8def prepare(9    tokenizer, class_token_index: int, rendered: str, length: int, max_candidates: int10) -> dict[str, np.ndarray]:11    full = tokenizer(rendered, truncation=False)12    if len(full["input_ids"]) > length:13        raise ValueError(f"Verdict prompt needs {len(full['input_ids'])} tokens; L{length} has no room")14    encoded = tokenizer(rendered, truncation=False, padding="max_length", max_length=length, return_tensors="np")15    ids = encoded["input_ids"].astype(np.int32)16    positions = np.flatnonzero(ids[0] == class_token_index)17    if len(positions) > max_candidates:18        raise ValueError("candidate markers exceed exported head capacity")19    markers = np.zeros((1, max_candidates, length), dtype=np.float32)20    for row, position in enumerate(positions):21        markers[0, row, position] = 1.022    return {23        "input_ids": ids,24        "attention_mask": encoded["attention_mask"].astype(np.int32),25        "class_marker_map": markers,26    }27