SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.12478 · 2026

Less Data, Faster Convergence: Goal-Driven Data Optimization for Multimodal Instruction Tuning

Yizhou Wang, Hai Ci, Rujie Wu, Haozhe Zhao

arXiv · PDF · Open in the Atlas

Code that ran

17 of 22 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; 7 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 9 failures of our instrument; 5 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/2603.12478.json · badge: /api/ran/2603.12478.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
rujiewu/GDO community
Papers with Code archive
17 ran · 5 no recorded run
FunctionStatusWhere it lives
compute_amm_and_diversity Ran
ran, honoured the contract we drafted
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("eea19ee437148fef")
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 tuple, 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 cv2
import numpy as np
import tempfile
import os

# Create temporary image files for testing
temp_dir = tempfile.mkdtemp()

# Create 3 simple test images (small RGB images)
image_paths = []
for i in range(3):
    # Create a simple 100x100 BGR image
    img = np.random.randint(0, 256, (100, 100, 3), dtype=np.uint8)
    img_path = os.path.join(temp_dir, f"frame_{i}.png")
    cv2.imwrite(img_path, img)
    image_paths.append(img_path)

ARGS = {
    "paths": image_paths,
    "max_frames": 2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
cv2, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple 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
is_oom_error Ran
ran, violated the contract we drafted
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("2cafe521fc07a22c")
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) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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

torch.manual_seed(0)

# The function takes an Exception and checks if it's an out-of-memory error
# We need to construct various Exception instances to test the function

# Test case 1: Exception with 'out of memory' in the message
exc_with_oom_msg = Exception("CUDA out of memory error")

# Test case 2: Regular exception without OOM message
exc_regular = Exception("Some other error")

# Test case 3: Exception with lowercase 'out of memory'
exc_lowercase_oom = RuntimeError("out of memory on device")

# We'll use the first one as the main test case
ARGS = {
    "exc": exc_with_oom_msg
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
iter_samples Ran
ran; our contract draft was wrong, not the code
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("420619a45dd0436c")
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 generator. 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 tempfile
import json
import os

# Create temporary files for onevision and video data
temp_dir = tempfile.mkdtemp()
onevision_path = os.path.join(temp_dir, 'onevision.jsonl')
video_path = os.path.join(temp_dir, 'video.jsonl')

# Write sample onevision data
onevision_data = [
    {
        'id': 'ov_001',
        'images': ['image1.jpg', 'image2.jpg'],
        'messages': [
            {'role': 'user', 'content': 'What is in this image?'},
            {'role': 'assistant', 'content': 'This is a test image.'},
            {'role': 'human', 'content': 'Tell me more <image>'},
            {'role': 'gpt', 'content': 'There is more detail here.'}
        ]
    },
    {
        'id': 'ov_002',
        'images': ['image3.jpg'],
        'messages': [
            {'role': 'from', 'value': 'Describe the scene'},
            {'role': 'gpt', 'value': 'A scenic landscape'}
        ]
    }
]

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

# Write sample video data
video_data = [
    {
        'id': 'vid_001',
        'video_id': 'video_uid_001',
        'images': ['frame1.jpg', 'frame2.jpg', 'frame3.jpg', 'frame4.jpg', 'frame5.jpg'],
        'messages': [
            {'role': 'user', 'content': 'What happens in this video? <video>'},
            {'role': 'assistant', 'content': 'The video shows action.'},
            {'role': 'human', 'content': 'Any other details?'},
            {'role': 'gpt', 'content': 'Yes, there are many details.'}
        ]
    },
    {
        'id': 'vid_002',
        'video_id': 'video_uid_002',
        'images': ['frame6.jpg', 'frame7.jpg'],
        'messages': [
            {'role': 'from', 'value': 'Analyze the video <video>'},
            {'role': 'gpt', 'value': 'The analysis is complete.'}
        ]
    }
]

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

# Define ARGS dictionary for iter_samples function
ARGS = {
    'onevision_path': onevision_path,
    'video_path': video_path,
    'target_frames': 3,
    'onevision_max_lines': 10,
    'video_max_lines': 10,
    'onevision_total_lines': 2,
    'video_total_lines': 2,
    'log_every': 1,
    'enable_tqdm': False,
    'print_parsing_log': False
}
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 generator — a generator output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a generator 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
jaccard_similarity Ran
ran, honoured the contract we drafted
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("cd6a7e9e2afeead9")
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) and returned float, 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 re
from typing import Dict, Iterator, List, Optional, Sequence, Tuple

WORD_RE = re.compile('[A-Za-z0-9_]+')

def tokenize(text: str) -> List[str]:
    return WORD_RE.findall((text or '').lower())

def jaccard_similarity(a: str, b: str) -> float:
    ta = set(tokenize(a))
    tb = set(tokenize(b))
    if not ta and (not tb):
        return 1.0
    inter = len(ta & tb)
    union = len(ta | tb)
    return float(inter / max(union, 1))

ARGS = {
    "a": "hello world test",
    "b": "hello python world"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 float, 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_pil_images Ran
ran, honoured the contract we drafted
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("08ef9725aaf447ef")
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) and returned list, ndim 4, dtype family int; 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 os
from PIL import Image
import numpy as np

# Create temporary directory with test images
temp_dir = tempfile.mkdtemp()

# Create a few test images
test_image_paths = []
for i in range(3):
    # Create a simple RGB image using PIL
    img_array = np.random.randint(0, 256, (10, 10, 3), dtype=np.uint8)
    img = Image.fromarray(img_array, mode='RGB')
    img_path = os.path.join(temp_dir, f'test_image_{i}.png')
    img.save(img_path)
    test_image_paths.append(img_path)

# Add a non-existent path to test exception handling
test_image_paths.append(os.path.join(temp_dir, 'nonexistent.png'))

ARGS = {
    'paths': test_image_paths
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
PIL, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 4, dtype family int — 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
make_texts Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("3e2ba7ffc4752aa5")
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 tuple, 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 torch.nn as nn
from transformers import AutoProcessor
from unittest.mock import Mock, MagicMock

torch.manual_seed(0)

# Create a minimal mock processor that implements apply_chat_template
class MockProcessor:
    def apply_chat_template(self, messages, tokenize=False, add_generation_prompt=False):
        # Simple mock implementation that returns a string representation
        if add_generation_prompt:
            return "<|im_start|>user\nquestion<|im_end|>\n<|im_start|>assistant\n"
        else:
            return "<|im_start|>user\nquestion<|im_end|>\n<|im_start|>assistant\nanswer<|im_end|>"

processor = MockProcessor()

ARGS = {
    'processor': processor,
    'question': 'What is in this image?',
    'answer': 'This is a test answer.',
    'num_images': 2,
    'use_images': True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple 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_text Ran
ran; our contract draft was wrong, not the code
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("db008612f143341f")
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) 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

torch.manual_seed(0)

# The function normalize_text takes a single parameter 'x' of type object
# It handles None, converts to string, strips whitespace, and joins with single spaces

# Test with various input types to ensure robustness
ARGS = {
    'x': '  hello   world  \n  test  '
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
parse_message_pairs Ran
ran; our contract draft was wrong, not the code
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("7b202711adf7f2cd")
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) and returned list, ndim 2, 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 Dict, List, Tuple

torch.manual_seed(0)

# Construct valid arguments for parse_message_pairs
ARGS = {
    'messages': [
        {
            'role': 'user',
            'content': 'Hello, what is <image> about?'
        },
        {
            'role': 'assistant',
            'content': 'This is a response about the image.'
        },
        {
            'role': 'human',
            'content': 'Tell me more about <video> content.'
        },
        {
            'role': 'gpt',
            'content': 'The video shows interesting things.'
        },
        {
            'role': 'user',
            'content': 'What about <audio>?'
        },
        {
            'role': 'assistant',
            'content': 'The audio contains speech and music.'
        }
    ]
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 2, 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
read_frames Ran
ran, honoured the contract we drafted
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("edaed27bb877f0f0")
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 4, dtype family int; 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 numpy as np
import tempfile
import os

# Create temporary image files for testing
temp_dir = tempfile.mkdtemp()

# Create a few dummy image files using numpy (simulating cv2.imread output)
# We'll create simple dummy files that cv2.imread can attempt to read
dummy_images = []
for i in range(5):
    # Create a simple 10x10 RGB image as numpy array
    img_array = np.random.randint(0, 255, (10, 10, 3), dtype=np.uint8)
    img_path = os.path.join(temp_dir, f"image_{i}.png")
    # Write using cv2 if available, otherwise just store paths
    try:
        import cv2
        cv2.imwrite(img_path, img_array)
        dummy_images.append(img_path)
    except Exception:
        dummy_images.append(img_path)

# Construct ARGS dict with valid arguments
ARGS = {
    "paths": dummy_images,  # Sequence of valid image file paths
    "max_frames": 3          # Target number of frames to sample
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
cv2, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 4, dtype family int — 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
sample_paths Ran
ran; our contract draft was wrong, not the code
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("c4fb4107dc8b9208")
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) 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

torch.manual_seed(0)

# Analyze the function:
# - paths: Sequence[str] - a sequence of strings
# - target: int - an integer target count
# The function filters out falsy values, converts to strings, and samples indices

ARGS = {
    "paths": ["path1", "path2", "path3", "path4", "path5"],
    "target": 3
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
self_consistency_score Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("e5992b29a0260e38")
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 float, 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 PIL import Image
from transformers import AutoProcessor
from unittest.mock import MagicMock, patch
from dataclasses import dataclass

torch.manual_seed(0)

@dataclass
class ModelBundle:
    processor: AutoProcessor
    model: object
    device: torch.device

# Create minimal mock processor
class MockProcessor:
    def apply_chat_template(self, messages, tokenize=False, add_generation_prompt=True):
        return "mock template text"
    
    def __call__(self, text=None, images=None, return_tensors=None, padding=None):
        # Return mock tensor dict with input_ids
        return {
            'input_ids': torch.randint(0, 100, (1, 10)),
            'attention_mask': torch.ones((1, 10), dtype=torch.long)
        }
    
    def batch_decode(self, token_ids, skip_special_tokens=False):
        return ["mock generated text response"]

# Create minimal mock model
class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
    
    def generate(self, input_ids=None, attention_mask=None, do_sample=False, 
                 top_p=None, temperature=None, max_new_tokens=None):
        batch_size = input_ids.shape[0]
        # Return generated tokens (batch_size, seq_len) - include prompt + new tokens
        return torch.randint(0, 100, (batch_size, 15))

# Create mock bundle
processor = MockProcessor()
model = MockModel()
device = torch.device('cpu')
bundle = ModelBundle(processor=processor, model=model, device=device)

# Create sample images
pil_images = [Image.new('RGB', (4, 3)), Image.new('RGB', (4, 3))]

ARGS = {
    'bundle': bundle,
    'question': 'What is in this image?',
    'pil_images': pil_images,
    'samples': 3,
    'max_new_tokens': 5
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, 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
strip_media_tokens Ran
ran; our contract draft was wrong, not the code
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("543eb8006243777b")
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) 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

MEDIA_TOKEN_RE = re.compile('<image>|<video>|<audio>')

def normalize_text(x: object) -> str:
    if x is None:
        return ''
    return ' '.join(str(x).strip().split())

def strip_media_tokens(text: str) -> str:
    return normalize_text(MEDIA_TOKEN_RE.sub(' ', text))

# Setup code
ARGS = {
    'text': 'This is a <image> test with <video> and <audio> tokens'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
temporal_regex_score Ran
ran, honoured the contract we drafted
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("098895353483bfe1")
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) and returned float, 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 re

TEMPORAL_RE = re.compile('\\b(before|after|then|while|during|first|next|finally|sequence|later|earlier|transition|happen)\\b', re.IGNORECASE)

def temporal_regex_score(question: str) -> float:
    return 1.0 if TEMPORAL_RE.search(question or '') else 0.0

# Test cases for the function
test_question_with_temporal = "What happens before the event?"
test_question_without_temporal = "What is the color of the sky?"

ARGS = {
    "question": test_question_with_temporal
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 float, 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
text_feature Ran
ran, honoured the contract we drafted
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("87c30108595426b1")
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) 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 re
from typing import Dict, Iterator, List, Optional, Sequence, Tuple
import numpy as np

WORD_RE = re.compile('[A-Za-z0-9_]+')

def tokenize(text: str) -> List[str]:
    return WORD_RE.findall((text or '').lower())

def text_feature(question: str, answer: str, bins: int=16) -> List[float]:
    vec = np.zeros((bins,), dtype=np.float32)
    toks = tokenize(question) + tokenize(answer)
    if not toks:
        return vec.tolist()
    for t in toks:
        idx = (hash(t) % bins + bins) % bins
        vec[idx] += 1.0
    vec /= float(np.linalg.norm(vec) + 1e-06)
    return vec.tolist()

ARGS = {
    "question": "What is machine learning?",
    "answer": "Machine learning is a subset of artificial intelligence.",
    "bins": 16
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
numpy — 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
tnc_llm Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("995b24926b03d47f")
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 float, 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 dataclasses import dataclass
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

torch.manual_seed(0)

class MockProcessor:
    def apply_chat_template(self, messages, tokenize=False, add_generation_prompt=False):
        return "mock prompt text"
    
    def __call__(self, text=None, return_tensors=None):
        batch_size = len(text) if text else 1
        return {
            'input_ids': torch.randint(0, 1000, (batch_size, 4)),
            'attention_mask': torch.ones(batch_size, 4)
        }
    
    def batch_decode(self, token_ids, skip_special_tokens=False):
        return ["temporal"]

class MockQwen3VL(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
    
    def generate(self, input_ids=None, attention_mask=None, do_sample=False, max_new_tokens=3, **kwargs):
        batch_size = input_ids.shape[0]
        seq_len = input_ids.shape[1]
        new_seq_len = seq_len + max_new_tokens
        return torch.randint(0, 1000, (batch_size, new_seq_len))

@dataclass
class ModelBundle:
    processor: AutoProcessor
    model: Qwen3VLForConditionalGeneration
    device: torch.device

mock_processor = MockProcessor()
mock_model = MockQwen3VL()
mock_device = torch.device('cpu')

bundle = ModelBundle(
    processor=mock_processor,
    model=mock_model,
    device=mock_device
)

question = "Does this video show motion?"

ARGS = {
    'bundle': bundle,
    'question': question
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, 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
tokenize Ran
ran; our contract draft was wrong, not the code
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("50bcada7381a2399")
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) 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 re
from typing import Dict, Iterator, List, Optional, Sequence, Tuple

WORD_RE = re.compile('[A-Za-z0-9_]+')

def tokenize(text: str) -> List[str]:
    return WORD_RE.findall((text or '').lower())

ARGS = {
    'text': 'Hello World 123 test_case'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
vision_feature Ran
ran, honoured the contract we drafted
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("271487e7d8776727")
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 os
import numpy as np
import cv2

# Create temporary directory with actual image files
temp_dir = tempfile.mkdtemp()

# Create a few simple test images
image_paths = []
for i in range(5):
    img = np.random.randint(0, 255, (10, 10, 3), dtype=np.uint8)
    path = os.path.join(temp_dir, f"test_image_{i}.png")
    cv2.imwrite(path, img)
    image_paths.append(path)

ARGS = {
    "paths": image_paths,
    "max_frames": 3
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
cv2, numpy — 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
ModelBundle Not yet run
unverified (harvested, no recorded run)
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("ba6e95dac46d63b8")
How it ran

no invocation recorded.

build_model Not yet run
unverified (harvested, no recorded run)
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("70149d03ea2ec0e9")
How it ran

no invocation recorded.

compute_loss Not yet run
unverified (harvested, no recorded run)
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("a71fb50b5b3bd84e")
How it ran

no invocation recorded.

main Not yet run
unverified (harvested, no recorded run)
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("3968b3aa5d9e95e8")
How it ran

no invocation recorded.

parse_args Not yet run
unverified (harvested, no recorded run)
rujiewu/GDO/gdo/extract_six_metrics.py
code served (permissive licence) · get_code("85315dad44b0f3f1")
How it ran

no invocation recorded.

Also present in other papers' code

Of 22 samples lifted from this paper's linked repositories, 22 carry a function identity and 0 do not. Of the 22 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 22 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 instruction tuning is often compute-inefficient because training budget is spread over large mixed image-video pools whose utility is highly uneven. We present Goal-Driven Data Optimization (GDO), a framework that computes six sample descriptors for each candidate and constructs optimized 1× training subsets for different goals. Under one fixed one-epoch Qwen3-VL-8B-Instruct training and evaluation recipe on 32 H20 GPUs, GDO uses far fewer training samples than Uni-10x while converging faster and reaching higher benchmark accuracy. Relative to the fixed 512k-sample Uni-10x baseline, GDO reaches the Uni-10x reference after 35.4k samples on MVBench, 26.6k on VideoMME, 27.3k on MLVU, and 34.7k on LVBench, while improving Accuracy by +1.38, +1.67, +3.08, and +0.84 pp, respectively. The gains are largest on MVBench and MLVU, while LVBench improves more modestly, consistent with its ultra-long-video setting and the mismatch between that benchmark and the short-video/image-dominant training pool. Across MinLoss, Diverse, Temp, and Temp+, stronger temporal pressure shifts the allocation toward video-centric supervision, with Temp+ giving the strongest overall profile. These results indicate that goal-driven data optimization improves sample efficiency and convergence under this fixed training contract. Code is available at https: //github.com/rujiewu/GDO.

For agents

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

get_harvested_code_for_paper("2603.12478")
get_code_for_paper("2603.12478")
have("2603.12478")

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

curl https://syntology.ai/api/ran/2603.12478.json

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

[![Syntology run record](https://syntology.ai/api/ran/2603.12478.svg)](https://syntology.ai/paper/2603.12478)

Connect an agent — have() is free.