scripts/run_reasoning_language_eval.py

from __future__ import annotations

import argparse
import datetime as dt
import json
import os
import random
import re
import time
import urllib.request
from pathlib import Path


SYSTEM_PROMPT = (
    "You are solving exact-answer reasoning problems. "
    "Follow the requested reasoning style exactly. "
    "Always end with a single final line formatted as ANSWER: <answer>."
)


CONDITIONS = {
    "direct": (
        "Solve the problem internally. "
        "Return only the final line. "
        "End with one line exactly in the form ANSWER: <answer>."
    ),
    "english": (
        "Reason step by step in plain English. "
        "Keep the reasoning concise but explicit. "
        "End with one line exactly in the form ANSWER: <answer>."
    ),
    "pseudocode": (
        "Reason in terse pseudocode with explicit running state updates. "
        "Prefer lines like `x = ...`, `total = ...`, `state -> ...`. "
        "End with one line exactly in the form ANSWER: <answer>."
    ),
    "lojban": (
        "Reason in Lojban before giving the answer. "
        "End with one line exactly in the form ANSWER: <answer>."
    ),
}


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser()
    parser.add_argument("--models", nargs="+", required=True)
    parser.add_argument("--conditions", nargs="+", required=True, choices=sorted(CONDITIONS))
    parser.add_argument("--output-dir", required=True)
    parser.add_argument("--task-count", type=int, default=18)
    parser.add_argument("--seed", type=int, default=20260406)
    parser.add_argument("--lojban-pilot-count", type=int, default=0)
    return parser.parse_args()


def make_arithmetic_task(rng: random.Random, task_id: str) -> dict[str, str]:
    a = rng.randint(14, 49)
    b = rng.randint(5, 17)
    c = rng.randint(3, 11)
    d = rng.randint(2, 9)
    e = rng.randint(6, 15)
    start = rng.randint(20, 80)
    answer = ((start + a + b) * c) - (d * e)
    prompt = (
        f"Start with {start}. Add {a}. Add {b}. Multiply the result by {c}. "
        f"Subtract {d} times {e}. What number do you get?"
    )
    return {"id": task_id, "kind": "arithmetic", "prompt": prompt, "answer": str(answer)}


def make_state_task(rng: random.Random, task_id: str) -> dict[str, str]:
    prompt = (
        "Run this exact program and report the final value of x - y.\n\n"
        f"x = {rng.randint(2, 8)}\n"
        f"y = {rng.randint(3, 9)}\n"
        f"z = {rng.randint(4, 10)}\n"
    )
    prompt_lines = prompt.splitlines()
    base_x = int(prompt_lines[2].split("=")[1].strip())
    base_y = int(prompt_lines[3].split("=")[1].strip())
    base_z = int(prompt_lines[4].split("=")[1].strip())
    x = base_x
    y = base_y
    z = base_z
    lines = [
        f"x = x + y",
        f"y = y * 2",
        f"z = z + x - 1",
        f"x = x * z",
        f"y = y + z",
    ]
    x = x + y
    y = y * 2
    z = z + x - 1
    x = x * z
    y = y + z
    answer = x - y
    prompt = (
        "Run this exact program and report the final value of x - y.\n\n"
        f"x = {base_x}\n"
        f"y = {base_y}\n"
        f"z = {base_z}\n"
        + "\n".join(lines)
    )
    return {"id": task_id, "kind": "state", "prompt": prompt, "answer": str(answer)}


def make_tasks(task_count: int, seed: int) -> list[dict[str, str]]:
    rng = random.Random(seed)
    tasks: list[dict[str, str]] = []
    for index in range(task_count):
        task_id = f"task-{index + 1:02d}"
        if index % 2 == 0:
            tasks.append(make_arithmetic_task(rng, task_id))
        else:
            tasks.append(make_state_task(rng, task_id))
    return tasks


def extract_answer(text: str) -> str:
    matches = re.findall(r"ANSWER:\s*(.+)", text)
    return matches[-1].strip() if matches else ""


def chat_completion(model: str, prompt: str) -> dict:
    payload = {
        "model": model,
        "temperature": 0,
        "messages": [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": prompt},
        ],
    }
    body = json.dumps(payload).encode("utf-8")
    request = urllib.request.Request(
        "https://api.openai.com/v1/chat/completions",
        data=body,
        headers={
            "Authorization": f"Bearer {os.environ['OPENAI_API_KEY']}",
            "Content-Type": "application/json",
        },
        method="POST",
    )
    with urllib.request.urlopen(request) as response:
        return json.loads(response.read().decode("utf-8"))


def main() -> None:
    args = parse_args()
    timestamp = dt.datetime.now(dt.timezone.utc).strftime("%Y%m%dT%H%M%SZ")
    output_dir = Path(args.output_dir) / timestamp
    output_dir.mkdir(parents=True, exist_ok=False)

    tasks = make_tasks(args.task_count, args.seed)
    (output_dir / "tasks.json").write_text(json.dumps(tasks, indent=2) + "\n")

    raw_path = output_dir / "raw.jsonl"
    rows: list[dict] = []

    for model in args.models:
        for condition in args.conditions:
            active_tasks = tasks
            if condition == "lojban" and args.lojban_pilot_count:
                active_tasks = tasks[: args.lojban_pilot_count]
            for task in active_tasks:
                prompt = f"{CONDITIONS[condition]}\n\nProblem:\n{task['prompt']}"
                started_at = time.time()
                response = chat_completion(model, prompt)
                ended_at = time.time()
                content = response["choices"][0]["message"]["content"]
                prediction = extract_answer(content)
                row = {
                    "model": model,
                    "condition": condition,
                    "task_id": task["id"],
                    "task_kind": task["kind"],
                    "gold_answer": task["answer"],
                    "predicted_answer": prediction,
                    "correct": prediction == task["answer"],
                    "latency_s": round(ended_at - started_at, 3),
                    "usage": response.get("usage", {}),
                    "response_id": response.get("id", ""),
                    "prompt": prompt,
                    "content": content,
                }
                rows.append(row)
                with raw_path.open("a", encoding="utf-8") as handle:
                    handle.write(json.dumps(row) + "\n")

    (output_dir / "results.json").write_text(json.dumps(rows, indent=2) + "\n")

    summaries: list[dict] = []
    for model in args.models:
        for condition in args.conditions:
            active_rows = [row for row in rows if row["model"] == model and row["condition"] == condition]
            correct = sum(1 for row in active_rows if row["correct"])
            total = len(active_rows)
            summaries.append(
                {
                    "model": model,
                    "condition": condition,
                    "correct": correct,
                    "total": total,
                    "accuracy": round(correct / total, 4),
                }
            )
    (output_dir / "summary.json").write_text(json.dumps(summaries, indent=2) + "\n")

    samples: list[str] = ["# Manual Samples", ""]
    for model in args.models:
        for condition in args.conditions:
            subset = [row for row in rows if row["model"] == model and row["condition"] == condition][:2]
            samples.append(f"## {model} / {condition}")
            samples.append("")
            for row in subset:
                samples.append(f"- {row['task_id']} `{row['task_kind']}` correct={row['correct']}")
                samples.append(f"  prompt: {row['prompt']}")
                samples.append(f"  output: {row['content']}")
                samples.append("")
    (output_dir / "samples.md").write_text("\n".join(samples) + "\n")

    print(output_dir)


if __name__ == "__main__":
    main()