scripts/chessqa_prompt_utils.py

"""
Utilities to prepare ChessQA-Benchmark prompts for evaluation.

The dataset questions contain placeholder tokens:
  - ``CONTEXT_PLACEHOLDER``: optionally replace with descriptive context.
  - ``FORMAT_EXAMPLE_PLACEHOLDER``: optionally replace with a sample answer format.

The helpers below make it easy to fill those placeholders consistently before
calling an LLM.
"""

from __future__ import annotations

import json
import re
from dataclasses import dataclass
from typing import Any, Dict, Mapping, Optional, Sequence, Tuple

import chess


@dataclass
class PromptConfig:
    """Configuration for prompt formatting."""

    add_context: bool = True
    """Inject piece arrangement and legal moves into the prompt when true."""

    format_example_index: int = 0
    """Pick which entry from ``format_examples`` to include (0-based)."""


def _clean_fen(fen: str) -> str:
    """Strip move annotations that follow the FEN with a pipe separator."""
    return fen.split("|", 1)[0].strip()


def get_context(fen: str) -> str:
    """
    Build a textual description for the provided FEN.

    Returns piece arrangement and all legal moves. When the FEN is invalid the
    function returns an empty string instead of raising.
    """
    fen_clean = _clean_fen(fen)
    try:
        board = chess.Board(fen_clean)
    except Exception:
        try:
            board = chess.Board(" ".join(fen_clean.split()))
        except Exception:
            return ""

    piece_names = {
        chess.PAWN: "Pawn",
        chess.KNIGHT: "Knight",
        chess.BISHOP: "Bishop",
        chess.ROOK: "Rook",
        chess.QUEEN: "Queen",
        chess.KING: "King",
    }

    pieces: Dict[str, list[str]] = {}
    for square in chess.SQUARES:
        piece = board.piece_at(square)
        if not piece:
            continue
        color = "White" if piece.color == chess.WHITE else "Black"
        key = f"{color} {piece_names[piece.piece_type]}"
        pieces.setdefault(key, []).append(chess.square_name(square))

    arrangement_parts = [f"{k}: {sorted(v)}" for k, v in pieces.items()]
    arrangement = ", ".join(arrangement_parts)
    legal_moves = ", ".join(sorted(move.uci() for move in board.legal_moves))
    return f"Piece arrangement: {arrangement}\nLegal moves: {legal_moves}\n\n"


def _select_format_example(format_examples: Optional[Sequence[str]], index: int) -> str:
    if not format_examples:
        return ""
    if index < 0 or index >= len(format_examples):
        index = 0
    return format_examples[index]


def _ensure_question_fields(row: Mapping[str, Any]) -> Tuple[str, Sequence[str]]:
    question = row.get("question", "")
    if not isinstance(question, str):
        raise TypeError("Expected 'question' to be a string")

    raw_examples = row.get("format_examples", [])
    if isinstance(raw_examples, str):
        try:
            raw_examples = json.loads(raw_examples)
        except json.JSONDecodeError:
            raw_examples = [raw_examples]

    if not isinstance(raw_examples, Sequence):
        raise TypeError("Expected 'format_examples' to be a sequence")
    return question, raw_examples


def format_prompt(
    row: Mapping[str, Any],
    config: Optional[PromptConfig] = None,
) -> str:
    """
    Replace placeholders in the dataset question.

    Parameters
    ----------
    row:
        A single dataset row (e.g., from the Hugging Face ``datasets`` library).
    config:
        Prompt configuration. When omitted defaults to ``PromptConfig()``.
    """
    config = config or PromptConfig()
    question, format_examples = _ensure_question_fields(row)

    formatted = question
    if config.add_context:
        fen = row.get("input", "")
        if isinstance(fen, str):
            context = get_context(fen)
        else:
            context = ""
        formatted = formatted.replace("CONTEXT_PLACEHOLDER", context)
    else:
        formatted = formatted.replace("CONTEXT_PLACEHOLDER", "")

    example_text = _select_format_example(format_examples, config.format_example_index)
    formatted = formatted.replace("FORMAT_EXAMPLE_PLACEHOLDER", example_text)
    return formatted


FINAL_ANSWER_PATTERN = re.compile(r"FINAL ANSWER:\\s*(.+?)(?:\\n|$)", re.IGNORECASE | re.DOTALL)


def extract_final_answer(text: str) -> Tuple[str, bool]:
    """
    Extract the trailing ``FINAL ANSWER:`` segment from a model response.

    Returns the extracted answer and a flag that indicates if the marker was
    found. Convenience helper for downstream evaluation scripts.
    """
    matches = list(FINAL_ANSWER_PATTERN.finditer(text or ""))
    if not matches:
        boxed = re.findall(r"[Tt]he\\s+final\\s+answer\\s+is\\s+\\$?\\\\boxed\\{([^}]+)\\}\\$?", text or "")
        if boxed:
            return boxed[-1].strip(), True
        return "", False

    answer = matches[-1].group(1).strip()
    answer = re.sub(r"^FINAL ANSWER:\\s*", "", answer, flags=re.IGNORECASE).strip()
    answer = answer.strip("*").strip()
    return answer, True


__all__ = [
    "PromptConfig",
    "extract_final_answer",
    "format_prompt",
    "get_context",
]