File size: 12,513 Bytes
bee2f63
 
 
 
 
 
 
 
 
86b8466
53f8f7c
90a12a7
bee2f63
 
 
 
 
 
90a12a7
bee2f63
 
 
 
 
 
90a12a7
bee2f63
 
 
 
90a12a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bee2f63
 
90a12a7
 
bee2f63
 
 
 
 
 
53f8f7c
 
 
90a12a7
 
bee2f63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
90a12a7
 
 
53f8f7c
 
 
 
 
 
90a12a7
 
53f8f7c
 
 
 
 
 
90a12a7
53f8f7c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
90a12a7
53f8f7c
 
 
 
 
 
 
 
86b8466
 
 
 
 
 
 
 
 
 
 
 
 
 
 
90a12a7
53f8f7c
bee2f63
 
 
90a12a7
bee2f63
 
 
 
 
53f8f7c
bee2f63
90a12a7
 
 
 
 
 
bee2f63
 
 
 
 
 
 
 
 
 
 
53f8f7c
bee2f63
 
86b8466
bee2f63
 
 
 
 
 
 
 
53f8f7c
bee2f63
 
 
 
 
 
 
 
 
53f8f7c
86b8466
53f8f7c
 
 
 
bee2f63
 
 
 
 
 
 
 
 
 
 
 
90a12a7
bee2f63
90a12a7
 
 
86b8466
53f8f7c
bee2f63
d4eadfe
 
 
e04e3db
 
 
bee2f63
 
 
90a12a7
bee2f63
 
 
 
90a12a7
 
 
bee2f63
 
90a12a7
 
 
 
 
 
bee2f63
 
 
 
90a12a7
bee2f63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
53f8f7c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bee2f63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
53f8f7c
 
 
 
 
bee2f63
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
53f8f7c
 
 
 
 
 
 
 
 
 
 
 
 
 
bee2f63
 
86b8466
bee2f63
 
 
 
 
 
 
 
 
 
 
 
 
53f8f7c
bee2f63
 
86b8466
bee2f63
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
"""
GAIA Test Bench - Local evaluation for your agent

Usage:
    uv run python test_bench.py                    # Run on 5 random questions
    uv run python test_bench.py --n 10             # Run on 10 random questions
    uv run python test_bench.py --level 1          # Run on level 1 only
    uv run python test_bench.py --level 1 --n 3    # Run on 3 level 1 questions
    uv run python test_bench.py --all              # Run on all validation questions
    uv run python test_bench.py --task-id 1234     # Run on specific task ID

    uv run python test_bench.py --type youtube,file,web,excel,pdf,image,audio,text-only
"""

import argparse
import json
import os
import re
from dataclasses import dataclass
from datetime import datetime
from pathlib import Path

from datasets import load_dataset
from dotenv import load_dotenv
from huggingface_hub import snapshot_download
from pydantic import BaseModel, Field

load_dotenv()


class AnnotatorMetadata(BaseModel):
    """Metadata from GAIA annotators about how to solve the task."""

    steps: str = Field(default="", alias="Steps")
    number_of_steps: str = Field(default="", alias="Number of steps")
    tools: str = Field(default="", alias="Tools")
    number_of_tools: str = Field(default="", alias="Number of tools")
    how_long: str = Field(default="", alias="How long did this take?")


class GaiaQuestion(BaseModel):
    """A single question from the GAIA dataset."""

    model_config = {"populate_by_name": True}

    task_id: str
    question: str = Field(alias="Question")
    level: int = Field(alias="Level")
    final_answer: str = Field(default="", alias="Final answer")
    file_name: str = ""
    file_path: str | None = None
    annotator: AnnotatorMetadata = Field(
        default_factory=AnnotatorMetadata, alias="Annotator Metadata"
    )


@dataclass
class TestResult:
    """Result of running an agent on a GAIA question."""

    task_id: str
    question: str
    expected: str
    actual: str
    correct: bool
    level: int
    file_name: str = ""
    file_path: str | None = None
    annotator_steps: str = ""
    annotator_tools: str = ""
    annotator_number_of_steps: str = ""


def normalize_answer(answer: str) -> str:
    """Normalize answer for comparison."""
    if not answer:
        return ""
    ans = answer.lower().strip()
    ans = re.sub(r"[.,;:!?]+$", "", ans)
    return ans


def answers_match(expected: str, actual: str) -> bool:
    """Check if answers match (with some flexibility)."""
    exp = normalize_answer(expected)
    act = normalize_answer(actual)

    if exp == act:
        return True

    if exp in act:
        return True

    try:
        exp_num = float(re.sub(r"[^\d.-]", "", exp))
        act_num = float(re.sub(r"[^\d.-]", "", act))
        if abs(exp_num - act_num) < 0.01:
            return True
    except (ValueError, TypeError):
        pass

    return False


def filter_by_type(
    questions: list[GaiaQuestion], task_type: str | None
) -> list[GaiaQuestion]:
    """Filter questions by task type."""
    if not task_type:
        return questions

    filtered = []
    for q in questions:
        question_text = q.question.lower()
        file_name = q.file_name.lower()

        match task_type:
            case "youtube":
                if "youtube.com" in question_text or "youtu.be" in question_text:
                    filtered.append(q)
            case "file":
                if q.file_name:
                    filtered.append(q)
            case "web":
                if "http://" in question_text or "https://" in question_text:
                    filtered.append(q)
            case "excel":
                if file_name.endswith((".xlsx", ".xls", ".csv")):
                    filtered.append(q)
            case "pdf":
                if file_name.endswith(".pdf"):
                    filtered.append(q)
            case "image":
                if file_name.endswith((".png", ".jpg", ".jpeg", ".gif", ".webp")):
                    filtered.append(q)
            case "audio":
                if file_name.endswith((".mp3", ".wav", ".m4a", ".flac", ".ogg")):
                    filtered.append(q)
            case "text-only":
                has_url = "http://" in question_text or "https://" in question_text
                has_file = bool(q.file_name)
                if not has_url and not has_file:
                    filtered.append(q)
            case _:
                filtered.append(q)

    return filtered


def filter_by_task_id(
    questions: list[GaiaQuestion], task_id: str | None
) -> list[GaiaQuestion]:
    """Filter questions by task id."""
    if not task_id:
        return questions

    filtered = []
    for q in questions:
        if q.task_id == task_id:
            filtered.append(q)

    return filtered


def load_gaia_data(level: int | None = None) -> list[GaiaQuestion]:
    """Load GAIA validation dataset with all metadata."""
    print("Downloading GAIA dataset...")
    data_dir = snapshot_download(repo_id="gaia-benchmark/GAIA", repo_type="dataset")

    questions: list[GaiaQuestion] = []

    levels = [level] if level else [1, 2, 3]
    for lvl in levels:
        try:
            dataset = load_dataset(data_dir, f"2023_level{lvl}", split="validation")

            for example in dataset:
                question = GaiaQuestion.model_validate(example)
                # Fix file_path to be absolute
                if question.file_path:
                    question.file_path = os.path.join(data_dir, question.file_path)
                questions.append(question)

        except Exception as e:
            print(f"Warning: Could not load level {lvl}: {e}")

    print(f"Loaded {len(questions)} questions")
    return questions


def run_test_bench(
    agent,
    n_questions: int | None = 5,
    level: int | None = None,
    task_type: str | None = None,
    run_all: bool = False,
    save_results: bool = True,
    task_id: str | None = None,
) -> list[TestResult]:
    """
    Run the test bench on the agent.

    Args:
        agent: The agent to test (callable that takes question string)
        n_questions: Number of questions to test (None for all)
        level: Filter by difficulty level (1, 2, or 3)
        task_type: Filter by task type (youtube, file, web, excel, pdf, image, audio, text-only)
        run_all: If True, run on all questions
        save_results: Save results to JSON file

    Returns:
        List of TestResult objects
    """
    import random

    questions = load_gaia_data(level=level)
    questions = filter_by_type(questions, task_type)
    questions = filter_by_task_id(questions, task_id)

    if not questions:
        print(f"No questions found for type '{task_type}'")
        return []

    if not run_all and n_questions and n_questions < len(questions):
        questions = random.sample(questions, n_questions)

    results: list[TestResult] = []
    correct_count = 0

    print(f"\n{'='*60}")
    print(f"Running {len(questions)} questions...")
    print("=" * 60)

    for i, q in enumerate(questions, 1):
        print(f"\n[{i}/{len(questions)}] Level {q.level}: {q.question[:80]}...")

        if q.file_name:
            print(f"  File: {q.file_name}")
        if q.annotator.tools:
            print(f"  Tools needed:\n{q.annotator.tools}")

        try:
            question_with_file = q.question
            if q.file_path:
                question_with_file += f"\n\nFile path: {q.file_path}"
            import asyncio

            actual = asyncio.run(agent(question_with_file))
        except Exception as e:
            actual = f"ERROR: {e}"

        is_correct = answers_match(q.final_answer, actual)
        if is_correct:
            correct_count += 1

        result = TestResult(
            task_id=q.task_id,
            question=q.question,
            expected=q.final_answer,
            actual=actual,
            correct=is_correct,
            level=q.level,
            file_name=q.file_name,
            file_path=q.file_path,
            annotator_steps=q.annotator.steps,
            annotator_tools=q.annotator.tools,
            annotator_number_of_steps=q.annotator.number_of_steps,
        )
        results.append(result)

        status = "CORRECT" if is_correct else "WRONG"
        print(f"  Expected: {q.final_answer}")
        print(f"  Got:      {actual[:100]}...")
        print(f"  Status:   {status}")

    # Summary
    print(f"\n{'='*60}")
    print("SUMMARY")
    print("=" * 60)
    print(f"Total:   {len(results)}")
    print(f"Correct: {correct_count}")
    print(f"Accuracy: {correct_count/len(results)*100:.1f}%")

    # Per-level breakdown
    for lvl in [1, 2, 3]:
        lvl_results = [r for r in results if r.level == lvl]
        if lvl_results:
            lvl_correct = sum(1 for r in lvl_results if r.correct)
            print(
                f"  Level {lvl}: {lvl_correct}/{len(lvl_results)} ({lvl_correct/len(lvl_results)*100:.1f}%)"
            )

    # Tool usage analysis
    print(f"\n{'='*60}")
    print("TOOL REQUIREMENTS ANALYSIS")
    print("=" * 60)
    tool_stats: dict[str, dict] = {}
    for r in results:
        tools = (
            r.annotator_tools
            if isinstance(r.annotator_tools, str)
            else ", ".join(r.annotator_tools)
        )
        if tools:
            if tools not in tool_stats:
                tool_stats[tools] = {"total": 0, "correct": 0}
            tool_stats[tools]["total"] += 1
            if r.correct:
                tool_stats[tools]["correct"] += 1

    for tools, stats in sorted(tool_stats.items(), key=lambda x: -x[1]["total"]):
        acc = stats["correct"] / stats["total"] * 100 if stats["total"] > 0 else 0
        print(f"  {tools}: {stats['correct']}/{stats['total']} ({acc:.0f}%)")

    # Save results
    if save_results:
        results_dir = Path("test_results")
        results_dir.mkdir(exist_ok=True)

        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
        results_file = results_dir / f"results_{timestamp}.json"

        with open(results_file, "w") as f:
            json.dump(
                {
                    "timestamp": timestamp,
                    "total": len(results),
                    "correct": correct_count,
                    "accuracy": correct_count / len(results),
                    "results": [
                        {
                            "task_id": r.task_id,
                            "level": r.level,
                            "question": r.question,
                            "expected": r.expected,
                            "actual": r.actual,
                            "correct": r.correct,
                            "file_name": r.file_name,
                            "file_path": r.file_path,
                            "annotator_steps": r.annotator_steps,
                            "annotator_tools": r.annotator_tools,
                            "annotator_number_of_steps": r.annotator_number_of_steps,
                        }
                        for r in results
                    ],
                },
                f,
                indent=2,
            )
        print(f"\nResults saved to: {results_file}")

    return results


def main():
    parser = argparse.ArgumentParser(description="GAIA Test Bench")
    parser.add_argument("--n", type=int, default=5, help="Number of questions to test")
    parser.add_argument("--level", type=int, choices=[1, 2, 3], help="Filter by level")
    parser.add_argument(
        "--type",
        choices=[
            "youtube",
            "file",
            "web",
            "excel",
            "pdf",
            "image",
            "audio",
            "text-only",
        ],
        help="Filter by task type",
    )
    parser.add_argument("--all", action="store_true", help="Run all questions")
    parser.add_argument("--no-save", action="store_true", help="Don't save results")
    parser.add_argument("--task-id", type=str, help="Run specific task ID")
    args = parser.parse_args()

    from agent import BasicAgent

    if callable(BasicAgent) and not hasattr(BasicAgent, "__call__"):
        agent = BasicAgent(question="")
    else:
        agent = BasicAgent

    run_test_bench(
        agent=agent,
        n_questions=args.n,
        level=args.level,
        task_type=args.type,
        run_all=args.all,
        save_results=not args.no_save,
        task_id=args.task_id,
    )


if __name__ == "__main__":
    main()