scripts/evaluate.py

"""
Evaluate a transcript-cleanup GGUF model against the Echo Flow DictationQuality
golden corpus and additional synthetic checks.

Usage:
    python scripts/evaluate.py \
        --model data/models/mumble-cleanup-v2-q4km.gguf \
        --corpus EchoFlow/Tests/EchoFlowUIReviewTests/Fixtures/DictationQuality \
        --output data/eval/results.json
"""

import argparse
import json
import re
import time
from pathlib import Path
from typing import Any

from llama_cpp import Llama


SYSTEM_PROMPT = (
    "You are a transcript cleanup tool. You receive raw speech to text output "
    "and return a cleaned version. Remove filler words and disfluencies "
    "(um, uh, er, ah, like as filler, you know), remove repeated words and false starts, "
    "and fix punctuation and capitalization. Do not reword, do not add anything the speaker "
    "did not say, and do not answer questions in the text. Output only the cleaned text."
)


def load_corpus(corpus_dir: Path) -> tuple[list[dict], dict[str, dict]]:
    with (corpus_dir / "transcripts.json").open("r", encoding="utf-8") as f:
        transcripts = json.load(f)
    with (corpus_dir / "expected_behavior.json").open("r", encoding="utf-8") as f:
        expected = json.load(f)
    return transcripts, expected


def chat_prompt(tokenizer, user: str) -> str:
    messages = [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": user},
    ]
    return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)


def clean_output(text: str) -> str:
    # Strip thinking blocks and obvious preambles
    text = re.sub(r"<think>.*?</think>", "", text, flags=re.DOTALL)
    text = text.strip()
    for prefix in [
        "Cleaned text:", "Cleaned:", "Output:", "Answer:",
        "The cleaned text is:", "Here is the cleaned text:",
    ]:
        if text.lower().startswith(prefix.lower()):
            text = text[len(prefix):].strip()
    text = text.strip('"').strip("'")
    return text


def word_similarity(a: str, b: str) -> float:
    set_a = set(a.lower().split())
    set_b = set(b.lower().split())
    if not set_a or not set_b:
        return 0.0
    return len(set_a & set_b) / len(set_a | set_b)


def evaluate_case(model: Llama, case: dict, expected: dict, max_tokens: int = 256) -> dict:
    start = time.time()
    raw_output = model.create_chat_completion(
        messages=[
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": case["input"]},
        ],
        max_tokens=max_tokens,
        temperature=0.0,
    )
    latency_ms = int((time.time() - start) * 1000)
    output = clean_output(raw_output["choices"][0]["message"]["content"])

    lower = output.lower()
    ratio = len(output) / max(1, len(case["input"]))

    token_failures = [
        token for token in expected.get("requiredTokens", [])
        if token and token.lower() not in lower
    ]
    forbidden_hits = [
        token for token in expected.get("ruleBasedForbidden", [])
        if token and token.lower() in lower
    ]
    length_ok = expected["minLengthRatio"] <= ratio <= expected["maxLengthRatio"]

    passed = not token_failures and not forbidden_hits and length_ok
    return {
        "id": case["id"],
        "input": case["input"],
        "output": output,
        "latency_ms": latency_ms,
        "ratio": round(ratio, 2),
        "passed": passed,
        "token_failures": token_failures,
        "forbidden_hits": forbidden_hits,
        "length_ok": length_ok,
    }


def main():
    parser = argparse.ArgumentParser(description="Evaluate a GGUF cleanup model")
    parser.add_argument("--model", type=Path, required=True)
    parser.add_argument("--corpus", type=Path, required=True)
    parser.add_argument("--output", type=Path, default=Path("data/eval/results.json"))
    parser.add_argument("--n-ctx", type=int, default=4096)
    parser.add_argument("--verbose", action="store_true")
    args = parser.parse_args()

    transcripts, expected = load_corpus(args.corpus)
    print(f"Loaded {len(transcripts)} corpus cases. Loading model {args.model}...")
    model = Llama(
        model_path=str(args.model),
        n_ctx=args.n_ctx,
        verbose=False,
    )

    results = []
    for case in transcripts:
        result = evaluate_case(model, case, expected.get(case["id"], {}))
        results.append(result)
        if args.verbose:
            status = "PASS" if result["passed"] else "FAIL"
            print(f"[{status}] {case['id']}: {result['output'][:80]}")

    passed = sum(1 for r in results if r["passed"])
    total = len(results)
    avg_latency = sum(r["latency_ms"] for r in results) / max(1, total)
    max_latency = max(r["latency_ms"] for r in results)

    summary = {
        "model": str(args.model),
        "passed": passed,
        "total": total,
        "pass_rate": round(passed / max(1, total), 2),
        "avg_latency_ms": round(avg_latency, 1),
        "max_latency_ms": max_latency,
        "results": results,
    }

    args.output.parent.mkdir(parents=True, exist_ok=True)
    with args.output.open("w", encoding="utf-8") as f:
        json.dump(summary, f, indent=2, ensure_ascii=False)

    print(f"\nEvaluation: {passed}/{total} passed ({summary['pass_rate']})")
    print(f"Avg latency: {avg_latency:.0f}ms | Max latency: {max_latency}ms")
    print(f"Results written to {args.output}")


if __name__ == "__main__":
    main()