SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2601.22162 · 2026

UniFinEval: Towards Unified Evaluation of Financial Multimodal Models across Text, Images and Videos

Qi Qi, Wei Zhang, Jun Han, Liwen Zhang, Zhenyu Wu, Junbo Zou, Xin Guo, Zheng Wei, Zhaowei Liu, Xiaoming Huang, Zhi Yang, Lingfeng Zeng, and 9 more

arXiv · PDF · Open in the Atlas

Code that ran

15 of 21 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 14 failures of our instrument; 6 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2601.22162.json · badge: /api/ran/2601.22162.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
aifinlab/UniFinEval canonical (classifier label)
classifier basis: census
15 ran · 6 no recorded run
FunctionStatusWhere it lives
build_process_value Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/result_utils.py
code served (permissive licence) · get_code("93a10bdfb8d89468")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re
from typing import Dict, Any

torch_seed = 0

ARGS = {
    'model_answer': 'The answer is 42. \\boxed{42}',
    'model_data': {
        'raw_response': {
            'choices': [
                {
                    'message': {
                        'reasoning': 'Let me think about this problem step by step. First, I need to understand what is being asked.',
                        'reasoning_content': None,
                        'reasoning_details': None
                    }
                }
            ]
        }
    }
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
clean_json_text Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/judge.py
code served (permissive licence) · get_code("a97875241ff90c42")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def clean_json_text(text: str) -> str:
    """
    从模型输出中提取 JSON 字符串
    兼容模型输出 ```json ... ``` 包裹的情况
    """
    text = text.strip()
    pattern = '```json\\s*(\\{.*?\\})\\s*```|(\\{.*\\})'
    match = re.search(pattern, text, re.DOTALL)
    if match:
        return match.group(1) or match.group(2)
    start = text.find('{')
    end = text.rfind('}')
    if start != -1 and end != -1:
        return text[start:end + 1]
    return text


# Test case 1: JSON wrapped in ```json markers
test_text_1 = """```json
{"key": "value", "number": 42}
"""

# Test case 2: Plain JSON without markers
test_text_2 = '{"name": "test", "data": [1, 2, 3]}'

# Test case 3: JSON with surrounding text
test_text_3 = 'Here is some text before {"result": "success"} and after'

# Test case 4: Whitespace handling
test_text_4 = """  
json
  {"status": "ok"}  
  
"""

ARGS = {
    "text": test_text_1
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
convert_to_standard_format Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/data_converter.py
code served (permissive licence) · get_code("e8760156d157dfd3")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json

# Construct a representative input item with various field formats
item = {
    'question_id': 'q_001',
    'image_id': 'img_001',
    'image_path': '/path/to/image1.jpg, /path/to/image2.jpg',
    'image_type': 'photo',
    'question_type': 'visual_qa',
    'original_image_path': '/original/path1.jpg;/original/path2.jpg',
    'image_urls': ['http://example.com/img1.jpg', 'http://example.com/img2.jpg'],
    'scenario': 'indoor',
    'capability': 'object_recognition',
    'difficulty': 'medium',
    'source': 'dataset_v1',
    'question': {
        'round1': 'What is in the image?',
        'round2': 'What color is it?'
    },
    'answer': {
        'round1': 'A cat',
        'round2': 'Orange'
    },
    'options': {
        'A': 'Dog',
        'B': 'Cat',
        'C': 'Bird',
        'D': 'Fish'
    },
    'qa_make_process': {
        'step1': 'Extract objects',
        'step2': 'Generate questions'
    },
    'extra_field_1': 'some_value',
    'extra_field_2': 42
}

ARGS = {
    'item': item
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
encode_image Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/model_api.py
code served (permissive licence) · get_code("625809a36981482c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import tempfile
from PIL import Image
import io

# Create a temporary test image
temp_dir = tempfile.mkdtemp()
test_image_path = os.path.join(temp_dir, "test_image.jpg")

# Create a simple test image (800x600 pixels)
img = Image.new('RGB', (800, 600), color=(73, 109, 137))
img.save(test_image_path, format='JPEG', quality=95)

# Set PIL_AVAILABLE flag for the function
PIL_AVAILABLE = True

# Define arguments for encode_image function
ARGS = {
    'image_path': test_image_path,
    'compress': False,
    'max_longest_side': 1600,
    'quality': 92
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
evaluate_single_item Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/evaluator.py
code served (permissive licence) · get_code("58ca03390a3351e2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import json
import logging
from typing import List, Dict, Any, Optional

# Configure logging
logging.basicConfig(level=logging.INFO)

# Mock the EVAL_CONFIG that the function references
EVAL_CONFIG = {
    'skip_missing_images': False
}

# Mock the normalize_question_type function that the function calls
def normalize_question_type(raw_type: str) -> str:
    """Mock function to normalize question types"""
    type_mapping = {
        '单选': '单选题',
        '多选': '多选题',
        '多轮单选': '多轮单选题',
    }
    return type_mapping.get(raw_type, raw_type)

# Create minimal valid arguments
item = {
    'question_id': 'q_001',
    'id': 'item_001',
    'question': 'What is 2+2?',
    'answer': 'The answer is 4',
    'options': {
        'A': '2',
        'B': '4',
        'C': '6',
        'D': '8'
    },
    'image_path': '',
    'image_urls': [],
    'question_type': '单选',
    'is_multi_round': False
}

enabled_models = [
    'gpt-4',
    'gpt-3.5-turbo',
    'claude-3-opus'
]

profiles = [
    'profile_expert',
    'profile_beginner',
    'profile_average'
]

workers = 2

ARGS = {
    'item': item,
    'enabled_models': enabled_models,
    'profiles': profiles,
    'workers': workers
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
format_options Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/prompts.py
code served (permissive licence) · get_code("15d98731e128c61f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from typing import Dict, Optional

# Test case 1: Normal dictionary with options
options_dict = {
    "A": "选项A内容",
    "B": "选项B内容",
    "C": "选项C内容"
}

# Test case 2: None (should return empty string)
options_none = None

# Test case 3: Empty dictionary (should return empty string)
options_empty = {}

# Test case 4: Dictionary with empty string values (should skip them)
options_with_empty = {
    "A": "选项A内容",
    "B": "",
    "C": "选项C内容"
}

# Primary test case - standard dictionary
ARGS = {
    "options": {
        "A": "选项A内容",
        "B": "选项B内容",
        "C": "选项C内容"
    }
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_image_format Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/model_api.py
code served (permissive licence) · get_code("571e6a28dad787d4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
from pathlib import Path

# Test with various image path formats
ARGS = {
    'image_path': '/path/to/image.png'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
is_answer_empty Ran
ran, violated the contract we drafted
aifinlab/UniFinEval/evaluate_py/result_utils.py
code served (permissive licence) · get_code("102ece5a1388ab6c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
from typing import Dict, Any

# Test case 1: Basic module2_item with valid answer
module2_item_valid = {
    'model': {
        'model_name': 'gpt-4',
        'answer': 'This is a valid answer'
    },
    'model1': {
        'model_name': 'gpt-3.5',
        'answer': 'Another valid answer'
    }
}

# Test case 2: module2_item with empty answer
module2_item_empty = {
    'model': {
        'model_name': 'gpt-4',
        'answer': ''
    }
}

# Test case 3: module2_item with dict answer (multi-round)
module2_item_dict_answer = {
    'model': {
        'model_name': 'gpt-4',
        'answer': {
            'round_1': 'Response 1',
            'round_2': 'Response 2'
        }
    }
}

# Test case 4: module2_item with None answer
module2_item_none = {
    'model': {
        'model_name': 'gpt-4',
        'answer': None
    }
}

# Test case 5: module2_item with whitespace-only answer
module2_item_whitespace = {
    'model': {
        'model_name': 'gpt-4',
        'answer': '   '
    }
}

# Test case 6: module2_item with multiple models
module2_item_multi = {
    'model': {
        'model_name': 'model-a',
        'answer': 'valid answer'
    },
    'model2': {
        'model_name': 'model-b',
        'answer': ''
    },
    'model3': {
        'model_name': 'model-c',
        'answer': {'round1': 'answer1', 'round2': ''}
    }
}

ARGS = {
    'module2_item': module2_item_valid,
    'model_key': 'model',
    'model_name': 'gpt-4'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_csv Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/data_loader.py
code served (permissive licence) · get_code("80031dd34b37c131")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import csv
from pathlib import Path

# Create a temporary CSV file for testing
temp_dir = tempfile.mkdtemp()
csv_file_path = Path(temp_dir) / "test_data.csv"

# Write test CSV data
csv_data = [
    {
        "id": "1",
        "image_path": "/path/to/image1.jpg;/path/to/image2.jpg",
        "image_urls": "http://example.com/img1.jpg;http://example.com/img2.jpg",
        "round": "1",
        "text": "sample text",
        "options": '{"A": "option_a", "B": "option_b"}',
    },
    {
        "id": "2",
        "image_path": "/path/to/image3.jpg",
        "image_urls": "http://example.com/img3.jpg",
        "round": "2",
        "text": "another sample",
        "option_key1": "value1",
        "option_key2": "value2",
    },
    {
        "id": "3",
        "image_path": "null",
        "image_urls": None,
        "round": "invalid",
        "text": "test with nulls",
        "options": None,
    },
]

with open(csv_file_path, 'w', newline='', encoding='utf-8') as f:
    fieldnames = ["id", "image_path", "image_urls", "round", "text", "options", "option_key1", "option_key2"]
    writer = csv.DictWriter(f, fieldnames=fieldnames)
    writer.writeheader()
    for row in csv_data:
        writer.writerow(row)

ARGS = {
    "file_path": csv_file_path,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_json Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/data_loader.py
code served (permissive licence) · get_code("d068ef448a266fb4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
from pathlib import Path

# Create a temporary directory for test files
temp_dir = tempfile.mkdtemp()

# Test case 1: JSON file with array structure
array_file = Path(temp_dir) / "array.json"
with open(array_file, 'w', encoding='utf-8') as f:
    json.dump([{"id": 1, "name": "item1"}, {"id": 2, "name": "item2"}], f)

# Test case 2: JSON file with dict containing 'items' field
items_file = Path(temp_dir) / "items.json"
with open(items_file, 'w', encoding='utf-8') as f:
    json.dump({"items": [{"id": 3, "name": "item3"}, {"id": 4, "name": "item4"}]}, f)

# Test case 3: JSON file with plain dict (no 'items' field)
dict_file = Path(temp_dir) / "dict.json"
with open(dict_file, 'w', encoding='utf-8') as f:
    json.dump({"id": 5, "name": "item5"}, f)

ARGS = {
    "file_path": array_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_jsonl Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/data_loader.py
code served (permissive licence) · get_code("fca995540b6041ae")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import tempfile
from pathlib import Path

# Create a temporary JSONL file for testing
temp_dir = tempfile.mkdtemp()
temp_file = Path(temp_dir) / "test_data.jsonl"

# Write some test data to the JSONL file
test_data = [
    {"id": 1, "text": "Hello world", "value": 42},
    {"id": 2, "text": "Another line", "value": 100},
    {"id": 3, "text": "Third entry", "nested": {"key": "value"}},
]

with open(temp_file, 'w', encoding='utf-8') as f:
    for item in test_data:
        f.write(json.dumps(item) + '\n')

# Define arguments for the function
ARGS = {
    "file_path": temp_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
merge_multi_round_items Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/data_converter.py
code served (permissive licence) · get_code("3a416d9bf9a752e6")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

# Setup code for merge_multi_round_items function

ARGS = {
    "items": [
        # Single-round item (no round field)
        {
            "id": "q1",
            "question_id": "q1",
            "question": "What is this?",
            "answer": "This is an object",
            "image_id": "img1",
            "image_urls": ["http://example.com/img1.jpg"],
            "question_type": "visual_qa",
            "scenario": "indoor",
            "capability": "object_detection",
            "difficulty": "easy",
            "source": "dataset_a"
        },
        # Multi-round items with same question_id
        {
            "id": "q2",
            "question_id": "q2",
            "round": 1,
            "question": "What color is it?",
            "answer": "It is red",
            "image_id": "img2",
            "image_urls": "http://example.com/img2.jpg",
        },
        {
            "id": "q2",
            "question_id": "q2",
            "round": 2,
            "question": "How many are there?",
            "answer": "There are three",
            "image_id": "img2",
            "image_urls": ["http://example.com/img2_alt.jpg"],
        },
        {
            "id": "q2",
            "question_id": "q2",
            "round": 3,
            "question": "What is the material?",
            "answer": "It is made of plastic",
            "image_id": "img2",
        },
        # Multi-round items with same id field only
        {
            "id": "q3",
            "round": 1,
            "question": "First question",
            "answer": "First answer",
            "image_id": "img3",
            "image_urls": "http://example.com/img3.jpg;http://example.com/img3_v2.jpg",
        },
        {
            "id": "q3",
            "round": 2,
            "question": "Second question",
            "answer": "Second answer",
            "image_id": "img3",
        },
        # Item with no id or question_id (should be skipped initially, added back at end)
        {
            "question": "Orphan question",
            "answer": "Orphan answer",
        }
    ]
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
normalize_question_type Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/prompts.py
code served (permissive licence) · get_code("f6d2ed65eee5b187")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

QUESTION_TYPE_MAPPING = {'single_choice': '单选题', 'multiple_choice': '多选题', 'true_false': '判断题', 'essay': '问答题', 'multi_round_single_choice': '多轮单选题', 'multi_round_essay': '多轮问答题', '单选题': '单选题', '多选题': '多选题', '判断题': '判断题', '问答题': '问答题', '多轮单选题': '多轮单选题', '多轮问答题': '多轮问答题', '单选': '单选题', '多选': '多选题', '判断': '判断题', '问答': '问答题', '多轮单选': '多轮单选题', '多轮问答': '多轮问答题', '单选题型': '单选题', '多选题型': '多选题', '判断题型': '判断题', '问答题型': '问答题', '单选题目': '单选题', '多选题目': '多选题', '判断题目': '判断题', '问答题目': '问答题', 'single': '单选题', 'multiple': '多选题', 'true_false_question': '判断题', 'essay_question': '问答题', 'multi_round_single': '多轮单选题', 'multi_round': '多轮问答题'}

def normalize_question_type(question_type: str) -> str:
    """
    标准化题型名称(统一转换为中文)
    
    Args:
        question_type: 题型名称(支持中英文)
        
    Returns:
        中文题型名称,如果不在支持列表中则返回原值
    """
    if not question_type:
        return ''
    question_type = question_type.strip()
    if question_type in QUESTION_TYPE_MAPPING:
        return QUESTION_TYPE_MAPPING[question_type]
    question_type_lower = question_type.lower()
    for key, value in QUESTION_TYPE_MAPPING.items():
        if key.lower() == question_type_lower:
            return value
    return question_type

ARGS = {
    'question_type': 'single_choice'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
sanitize_messages_for_log Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/logger.py
code served (permissive licence) · get_code("b28da1e86c8bd6f5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from typing import List, Dict, Any, Optional

torch.manual_seed(0)

# Construct valid arguments for sanitize_messages_for_log

# Test case 1: messages with base64 image data and image_paths provided
messages_with_base64 = [
    {
        'role': 'user',
        'content': [
            {'type': 'text', 'text': 'What is in this image?'},
            {'type': 'image_url', 'image_url': {'url': 'data:image/jpeg;base64,/9j/4AAQSkZJRgABAQEA...'}}
        ]
    },
    {
        'role': 'assistant',
        'content': [
            {'type': 'text', 'text': 'This is an image.'},
            {'type': 'image_url', 'image_url': {'url': 'data:image/png;base64,iVBORw0KGgoAAAANSUhEUg...'}}
        ]
    }
]

image_paths_list = ['/path/to/image1.jpg', '/path/to/image2.png']

# Test case 2: messages with regular URLs (not base64)
messages_with_urls = [
    {
        'role': 'user',
        'content': [
            {'type': 'text', 'text': 'Check this URL'},
            {'type': 'image_url', 'image_url': {'url': 'https://example.com/image.jpg'}}
        ]
    }
]

# Test case 3: empty messages
empty_messages = []

# Test case 4: messages with string content (not list)
messages_string_content = [
    {
        'role': 'user',
        'content': 'This is a simple text message'
    }
]

# Test case 5: messages with base64 but no image_paths
messages_base64_no_paths = [
    {
        'role': 'user',
        'content': [
            {'type': 'text', 'text': 'Image below'},
            {'type': 'image_url', 'image_url': {'url': 'data:image/webp;base64,UklGRiYA...'}}
        ]
    }
]

ARGS = {
    'messages': messages_with_base64,
    'image_paths': image_paths_list
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
strip_boxed_content Ran
ran; our contract draft was wrong, not the code
aifinlab/UniFinEval/evaluate_py/result_utils.py
code served (permissive licence) · get_code("2a0691b3aab83548")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def strip_boxed_content(text: str) -> str:
    """
    移除文本中的 \\boxed{...} 片段,只保留思考/分析部分。
    
    Args:
        text: 原始文本
        
    Returns:
        清理后的文本
    """
    if not text:
        return ''
    cleaned = re.sub('\\\\\\\\boxed\\{.*?\\}', '', text, flags=re.DOTALL)
    cleaned = re.sub('\\\\boxed\\{.*?\\}', '', cleaned, flags=re.DOTALL)
    return cleaned.strip()


ARGS = {
    "text": "This is some thinking \\boxed{answer1} and more text \\\\boxed{answer2} with analysis"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
calculate_output_statistics Not yet run
unverified (harvested, no recorded run)
aifinlab/UniFinEval/evaluate_py/statistics.py
code served (permissive licence) · get_code("df38989e5f169bd9")
How it ran

no invocation recorded.

calculate_statistics Not yet run
unverified (harvested, no recorded run)
aifinlab/UniFinEval/evaluate_py/statistics.py
code served (permissive licence) · get_code("197585381b7bb3fd")
How it ran

no invocation recorded.

compress_image Not yet run
unverified (harvested, no recorded run)
aifinlab/UniFinEval/evaluate_py/model_api.py
code served (permissive licence) · get_code("6eb801bbd3095e75")
How it ran

no invocation recorded.

convert_batch Not yet run
unverified (harvested, no recorded run)
aifinlab/UniFinEval/evaluate_py/data_converter.py
code served (permissive licence) · get_code("ff8af687b81c85d6")
How it ran

no invocation recorded.

convert_to_module2_format Not yet run
unverified (harvested, no recorded run)
aifinlab/UniFinEval/evaluate_py/result_converter.py
code served (permissive licence) · get_code("5a452f073e882c48")
How it ran

no invocation recorded.

get_prompt Not yet run
unverified (harvested, no recorded run)
aifinlab/UniFinEval/evaluate_py/prompts.py
code served (permissive licence) · get_code("eb48ceb1095ac012")
How it ran

no invocation recorded.

Also present in other papers' code

Of 21 samples lifted from this paper's linked repositories, 21 carry a function identity and 0 do not. Of the 21 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 21 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Multimodal large language models are playing an increasingly significant role in empowering the financial domain, however, the challenges they face, such as multimodal and high-density information and cross-modal multi-hop reasoning, go beyond the evaluation scope of existing multimodal benchmarks. To address this gap, we propose UniFinEval, the first unified multimodal benchmark designed for high-information-density financial environments, covering text, images, and videos. UniFinEval systematically constructs five core financial scenarios grounded in real-world financial systems: Financial Statement Auditing, Company Fundamental Reasoning, Industry Trend Insights, Financial Risk Sensing, and Asset Allocation Analysis. We manually construct a high-quality dataset consisting of 3,767 question-answer pairs in both chinese and english and systematically evaluate 10 mainstream MLLMs under Zero-Shot and CoT settings. Results show that Gemini-3-pro-preview achieves the best overall performance, yet still exhibits a substantial gap compared to financial experts. Further error analysis reveals systematic deficiencies in current models. UniFinEval aims to provide a systematic assessment of MLLMs' capabilities in fine-grained, high-information-density financial environments, thereby enhancing the robustness of MLLMs applications in real-world financial scenarios. Data and code are available at https://github.com/aifinlab/UniFinEval.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2601.22162")
get_code_for_paper("2601.22162")
have("2601.22162")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2601.22162.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2601.22162.svg)](https://syntology.ai/paper/2601.22162)

Connect an agent — have() is free.