benchpress_eval/data.py

from __future__ import annotations

import json
from pathlib import Path
from typing import Any

import numpy as np
import pandas as pd
from PIL import Image

REQUIRED_COLUMNS = (
    "sample_id",
    "edit_type",
    "before",
    "after",
    "instruction_pos",
    "instruction_neg_list",
    "instruction_neg_types",
)

PATH_COLUMNS = ("before", "after")


def load_dataset(dataset_root: str | Path) -> pd.DataFrame:
    dataset_root = Path(dataset_root)
    parquet_path = dataset_root / "benchmark.parquet"
    if not parquet_path.exists():
        raise FileNotFoundError(f"Missing dataset parquet: {parquet_path}")
    return pd.read_parquet(parquet_path)


def as_list(value: Any) -> list[Any]:
    if value is None:
        return []
    if isinstance(value, float) and np.isnan(value):
        return []
    if isinstance(value, np.ndarray):
        return value.tolist()
    if isinstance(value, (list, tuple)):
        return list(value)
    return [value]


def parse_metadata(value: Any) -> dict[str, Any]:
    if isinstance(value, dict):
        return value
    if isinstance(value, str):
        try:
            parsed = json.loads(value)
            return parsed if isinstance(parsed, dict) else {}
        except json.JSONDecodeError:
            return {}
    return {}


def expand_triplets(df: pd.DataFrame) -> pd.DataFrame:
    """Expand edit-level rows into positive/negative verification triplets."""
    rows: list[dict[str, Any]] = []
    for row_index, row in df.reset_index(drop=True).iterrows():
        sample_id = row.get("sample_id", f"row-{row_index:05d}")
        base = {
            "sample_id": sample_id,
            "parquet_row_index": int(row.get("parquet_row_index", row_index)),
            "edit_type": row["edit_type"],
            "before": row["before"],
            "after": row["after"],
        }
        rows.append(
            {
                **base,
                "instruction": row["instruction_pos"],
                "label": 1,
                "ground_truth": True,
                "example_type": "positive",
                "negative_type": "positive",
                "negative_index": -1,
            }
        )

        negs = as_list(row["instruction_neg_list"])
        neg_types = as_list(row["instruction_neg_types"])
        if len(negs) != len(neg_types):
            raise ValueError(f"Negative instruction/type length mismatch for {sample_id}")
        for neg_index, (instruction, negative_type) in enumerate(zip(negs, neg_types)):
            rows.append(
                {
                    **base,
                    "instruction": instruction,
                    "label": 0,
                    "ground_truth": False,
                    "example_type": "negative",
                    "negative_type": str(negative_type),
                    "negative_index": neg_index,
                }
            )
    return pd.DataFrame(rows)


def _is_portable_relative_path(value: Any) -> bool:
    if value is None:
        return False
    path = Path(str(value))
    return not path.is_absolute() and ".." not in path.parts


def validate_dataset(dataset_root: str | Path, strict_core: bool = True) -> dict[str, Any]:
    """Validate the public Hugging Face dataset layout."""
    dataset_root = Path(dataset_root)
    df = load_dataset(dataset_root)
    report: dict[str, Any] = {
        "row_count": int(len(df)),
        "required_missing": [col for col in REQUIRED_COLUMNS if col not in df.columns],
        "edit_type_counts": df["edit_type"].value_counts().sort_index().to_dict()
        if "edit_type" in df.columns
        else {},
    }

    before_exists = df["before"].map(lambda p: (dataset_root / str(p)).exists())
    after_exists = df["after"].map(lambda p: (dataset_root / str(p)).exists())
    report["before_paths_resolvable"] = int(before_exists.sum())
    report["after_paths_resolvable"] = int(after_exists.sum())

    length_matches = []
    for _, row in df.iterrows():
        length_matches.append(
            len(as_list(row["instruction_neg_list"])) == len(as_list(row["instruction_neg_types"]))
        )
    report["negative_instruction_lengths_match"] = int(sum(length_matches))

    non_portable_paths: dict[str, int] = {}
    for col in PATH_COLUMNS:
        if col in df.columns:
            bad_count = int((~df[col].map(_is_portable_relative_path)).sum())
            if bad_count:
                non_portable_paths[col] = bad_count
    report["non_portable_path_columns"] = non_portable_paths

    dimensions: dict[str, int] = {}
    for path_text in pd.concat([df["before"], df["after"]]).head(100):
        path = dataset_root / str(path_text)
        if path.exists():
            with Image.open(path) as image:
                key = f"{image.width}x{image.height}"
                dimensions[key] = dimensions.get(key, 0) + 1
    report["sampled_image_dimension_counts"] = dimensions

    checks = [
        not report["required_missing"],
        report["before_paths_resolvable"] == len(df),
        report["after_paths_resolvable"] == len(df),
        report["negative_instruction_lengths_match"] == len(df),
        not report["non_portable_path_columns"],
    ]
    if strict_core:
        checks.extend(
            [
                len(df) == 1500,
                set(report["edit_type_counts"].values()) == {150},
            ]
        )
    report["passed"] = bool(all(checks))
    return report