SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2410.10034 · 2024

TULIP: Token-length Upgraded CLIP

arXiv · PDF · Open in the Atlas

Code that ran

13 of 20 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; 2 honoured the contract we drafted, 1 violated it and 1 has no contract verdict; 9 failures of our instrument; 7 no recorded run); 1 of them was also run on the host, not in the jail, by the peer-agreement lane that set its level, measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; 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/2410.10034.json · badge: /api/ran/2410.10034.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
ivonajdenkoska/tulip canonical (classifier label)
classifier basis: census
13 ran · 7 no recorded run
FunctionStatusWhere it lives
natural_key Ran
ran; our contract draft was wrong, not the code
ivonajdenkoska/tulip/eval_tulip.py
code served (permissive licence) · get_code("a677092c48744688")
How it ran

Ran in Syntology's jail 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. The comparison behind this sample's level is a different run: the peer-agreement lane ran it again on the host, not in the jail (report dated 2026-09-05T01:20:25Z). 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

def natural_key(string_):
    """See http://www.codinghorror.com/blog/archives/001018.html"""
    return [int(s) if s.isdigit() else s for s in re.split('(\\d+)', string_.lower())]

ARGS = {
    "string_": "file_123_name_456.txt"
}
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
collate_fn Ran
ran; our contract draft was wrong, not the code
ivonajdenkoska/tulip/eval/dci_long.py
code served (permissive licence) · get_code("09c78a125c9f58ae")
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. 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

torch.manual_seed(0)

# Create sample data: batch is a list of (image, caption) tuples
# collate_fn expects images to be stackable tensors and captions to be strings

batch = [
    (torch.randn(3, 4, 5), "a cat sitting on a mat"),
    (torch.randn(3, 4, 5), "a dog playing fetch"),
]

ARGS = {
    "batch": batch
}
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 tuple — a tuple output has no array shape of its own; none is 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_custom_text_state_dict Ran
ran, violated the contract we drafted
ivonajdenkoska/tulip/open_clip/model.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from deep-real/dcp, is pointer only (licence: NONE) · get_code("f19962ebb134b3d7")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from deep-real/dcp, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_accuracy_t2i Ran
ran; our contract draft was wrong, not the code
ivonajdenkoska/tulip/eval/flickr30.py
code served (permissive licence) · get_code("385a40328778e8cb")
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. 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
import torch.nn as nn

torch.manual_seed(0)

# Create small feature tensors
# The function expects:
# - text_feature: normalized text embeddings, shape (batch_size, feature_dim)
# - image_feature: normalized image embeddings, shape (batch_size, feature_dim)
# - k: number of top-k results to check
#
# The function divides batch_size by 5 to get true_index, so we need batch_size divisible by 5
# Let's use batch_size=10 (divisible by 5) and feature_dim=4

batch_size = 10
feature_dim = 4

# Create random feature vectors (will be normalized by the function)
text_feature = torch.randn(batch_size, feature_dim)
image_feature = torch.randn(batch_size, feature_dim)

# k is the top-k parameter for accuracy checking
k = 2

ARGS = {
    "text_feature": text_feature,
    "image_feature": image_feature,
    "k": k
}
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 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
get_cast_dtype Ran
ran; our contract draft was wrong, not the code
ivonajdenkoska/tulip/open_clip/model.py
code served (permissive licence) · get_code("dcd422d66b0581d8")
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 dtype. 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

torch.manual_seed(0)

ARGS = {
    "precision": "bf16"
}
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 dtype — a dtype output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dtype 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_clip_metrics Ran
ran; our contract draft was wrong, not the code
ivonajdenkoska/tulip/eval/dci_long.py
code served (permissive licence) · get_code("0b9ff430cf687b70")
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 torch
import numpy as np

torch.manual_seed(0)

# Create small tensors for image and text features
# image_features: (batch_size, embedding_dim)
# text_features: (batch_size, embedding_dim)
batch_size = 2
embedding_dim = 4

image_features = torch.randn(batch_size, embedding_dim)
text_features = torch.randn(batch_size, embedding_dim)

# logit_scale is a scalar tensor that scales the similarity scores
logit_scale = torch.tensor(1.0)

ARGS = {
    'image_features': image_features,
    'text_features': text_features,
    'logit_scale': logit_scale,
}
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 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
get_image_feature Ran
ran, honoured the contract we drafted
ivonajdenkoska/tulip/eval/flickr30.py
code served (permissive licence) · get_code("0434c3049ad999d2")
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 Tensor, ndim 2, dtype family float, shape (2, 4) 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
import torch.nn as nn
from PIL import Image
import io
import tempfile
import os

torch.manual_seed(0)

# Create a minimal model with encode_image method
class MockVisionModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(3, 4)  # 3 channels to 4-dim feature
    
    def encode_image(self, image):
        # Expects image tensor of shape (batch, 3, height, width)
        # Returns (batch, 4) feature vector
        batch_size = image.shape[0]
        return self.linear(image.mean(dim=[2, 3]))  # Global average pool then linear

model = MockVisionModel()

# Create a minimal processor that converts PIL Image to tensor
class MockProcessor:
    def __call__(self, image):
        # Convert PIL Image to tensor of shape (3, height, width)
        img_array = torch.tensor([[[1.0, 2.0], [3.0, 4.0]],
                                   [[5.0, 6.0], [7.0, 8.0]],
                                   [[9.0, 10.0], [11.0, 12.0]]], dtype=torch.float32)
        return img_array

processor = MockProcessor()

# Create temporary directory structure with mock data
temp_dir = tempfile.mkdtemp()
datasets_dir = os.path.join(temp_dir, 'Datasets', 'flickr', 'flickr30k_images', 'flickr30k_images')
os.makedirs(datasets_dir, exist_ok=True)

# Create a mock CSV file
csv_path = os.path.join(temp_dir, 'Datasets', 'flickr', 'results.csv')
os.makedirs(os.path.dirname(csv_path), exist_ok=True)

# Create mock image files
mock_img = Image.new('RGB', (64, 64), color='red')
img_names = [f'image_{i}.jpg' for i in range(10)]
for img_name in img_names:
    mock_img.save(os.path.join(datasets_dir, img_name))

# Create results.csv with proper format
with open(csv_path, 'w') as f:
    f.write('image_id|caption\n')
    for i, img_name in enumerate(img_names):
        f.write(f'{img_name}|caption_{i}\n')

data_path = temp_dir + '/'

ARGS = {
    'model': model,
    'processor': processor,
    'data_path': data_path,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) 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
get_input_dtype Ran
ran; our contract draft was wrong, not the code
ivonajdenkoska/tulip/open_clip/model.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from deep-real/dcp, is pointer only (licence: NONE) · get_code("b476c8cfbf0f1f47")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from deep-real/dcp, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_latest_checkpoint Ran
ran; our contract draft was wrong, not the code
ivonajdenkoska/tulip/eval_tulip.py
code served (permissive licence) · get_code("93448c613c050973")
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 tempfile
import os
import torch

# Create a temporary directory with some dummy checkpoint files
temp_dir = tempfile.mkdtemp()

# Create some dummy .pt files with names that will be sorted naturally
checkpoint_files = [
    os.path.join(temp_dir, "checkpoint_1.pt"),
    os.path.join(temp_dir, "checkpoint_10.pt"),
    os.path.join(temp_dir, "checkpoint_2.pt"),
]

for checkpoint_file in checkpoint_files:
    torch.save({"dummy": "state"}, checkpoint_file)

# The function expects a path pattern that can be globbed
# It uses glob.glob(path + '**/*.pt', recursive=True)
# So we need to provide a path that ends with a separator
path_arg = temp_dir + os.sep

ARGS = {
    "path": path_arg,
}
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 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_text_feature Ran
ran, honoured the contract we drafted
ivonajdenkoska/tulip/eval/flickr30.py
code served (permissive licence) · get_code("5be654577f88529b")
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 Tensor, ndim 2, dtype family float, shape (4000, 4) 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
import torch.nn as nn
import pandas as pd
import tempfile
import os

torch.manual_seed(0)

# Create a minimal processor with tokenizer
class MinimalTokenizer:
    def __call__(self, text_list):
        # Return a mock tokenized object with a to() method
        return MockTokenized()

class MockTokenized:
    def to(self, device):
        return self

# Create a minimal processor
class MinimalProcessor:
    def __init__(self):
        self.tokenizer = MinimalTokenizer()

# Create a minimal distilled model
class MinimalDistilledModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(1, 4)
    
    def encode_text(self, text_input):
        # Return a small tensor (batch_size, embedding_dim)
        # We'll return a tensor of shape (batch_size, 4)
        batch_size = 2  # Small batch
        return torch.randn(batch_size, 4)

# Create a minimal full model (not used but required as parameter)
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(1, 1)

# Create a temporary CSV file with test data
def create_test_csv():
    temp_dir = tempfile.mkdtemp()
    os.makedirs(os.path.join(temp_dir, 'Datasets', 'flickr'), exist_ok=True)
    
    # Create a small CSV with text data
    data = {
        'image_name': [f'img_{i}.jpg' for i in range(100)],
        'comment_number': list(range(100)),
        'comment': [f'This is comment number {i}' for i in range(100)]
    }
    df = pd.DataFrame(data)
    csv_path = os.path.join(temp_dir, 'Datasets', 'flickr', 'results.csv')
    df.to_csv(csv_path, sep='|', index=False)
    
    return temp_dir + '/'

# Setup arguments
model = MinimalModel()
distilled_model = MinimalDistilledModel()
processor = MinimalProcessor()
data_path = create_test_csv()

ARGS = {
    'model': model,
    'distilled_model': distilled_model,
    'processor': processor,
    'data_path': data_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
pandas, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (4000, 4) — shape (4000, 4) 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
prepare_inputs_for_generation Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
ivonajdenkoska/tulip/open_clip/coca_model.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from WHU-USI3DV/TreeCLS, is pointer only (licence: NONE) · get_code("fb651d0a97fd4d3f")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from WHU-USI3DV/TreeCLS, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

register_pooler Ran
ran on a synthesized input
ivonajdenkoska/tulip/open_clip/hf_model.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from WHU-USI3DV/TreeCLS, is pointer only (licence: NONE) · get_code("2a377da4a76a2d44")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from WHU-USI3DV/TreeCLS, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

run_urban1k_openclip Ran
ran; our contract draft was wrong, not the code
ivonajdenkoska/tulip/eval/urban1k.py
code served (permissive licence) · get_code("287dbd8a0930623d")
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. 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 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
import torch.utils.data as data
from unittest.mock import Mock
import tempfile
import os

torch.manual_seed(0)

# Create temporary directories to satisfy the dataset
temp_dir = tempfile.mkdtemp()
image_dir = os.path.join(temp_dir, 'Datasets/Urban1k/image')
caption_dir = os.path.join(temp_dir, 'Datasets/Urban1k/caption')
os.makedirs(image_dir, exist_ok=True)
os.makedirs(caption_dir, exist_ok=True)

# Create minimal test files
from PIL import Image as PILImage
import numpy as np

for i in range(2):
    # Create dummy images
    img_array = np.random.randint(0, 256, (8, 8, 3), dtype=np.uint8)
    img = PILImage.fromarray(img_array)
    img.save(os.path.join(image_dir, f'image_{i}.jpg'))
    
    # Create dummy captions
    with open(os.path.join(caption_dir, f'image_{i}.txt'), 'w') as f:
        f.write(f'This is a caption for image {i}')

# Create mock model with encode_image method
class MockModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(3, 4)
    
    def encode_image(self, images):
        # images shape: (batch_size, 3, H, W) - mock expects (batch, 3)
        batch_size = images.shape[0]
        return torch.randn(batch_size, 4)

# Create mock distilled model with encode_text method
class MockDistilledModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(10, 4)
    
    def encode_text(self, text_encoded):
        # Expects tokenized text, should return features
        batch_size = len(text_encoded['input_ids'])
        return torch.randn(batch_size, 4)

# Create mock processor
class MockProcessor:
    def __init__(self):
        self.tokenizer = MockTokenizer()
    
    def __call__(self, img):
        # Returns a tensor representation of the image
        return torch.randn(3, 8, 8)

class MockTokenizer:
    def __call__(self, captions):
        # Takes list of captions, returns dict with input_ids
        if isinstance(captions, str):
            captions = [captions]
        return MockTokenizedText(len(captions))

class MockTokenizedText(dict):
    def __init__(self, batch_size):
        super().__init__()
        self['input_ids'] = torch.randint(0, 100, (batch_size, 10))
    
    def to(self, device):
        return self

# Build arguments
model = MockModel()
distilled_model = MockDistilledModel()
processor = MockProcessor()
data_path = temp_dir + '/'

ARGS = {
    'model': model,
    'distilled_model': distilled_model,
    'processor': processor,
    'data_path': data_path,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, torch, tqdm — 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
gather_features Not yet run
unverified (harvested, no recorded run)
ivonajdenkoska/tulip/open_clip/loss.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from lxa9867/imagefolder, is pointer only (licence: NOASSERTION) · get_code("ddcbd45e940484ee")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from lxa9867/imagefolder, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

neighbour_exchange Not yet run
unverified (harvested, no recorded run)
ivonajdenkoska/tulip/open_clip/loss.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from lxa9867/imagefolder, is pointer only (licence: NOASSERTION) · get_code("e332856e3c2fc814")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from lxa9867/imagefolder, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

neighbour_exchange_bidir Not yet run
unverified (harvested, no recorded run)
ivonajdenkoska/tulip/open_clip/loss.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from lxa9867/imagefolder, is pointer only (licence: NOASSERTION) · get_code("5b1fd364afcf3c05")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from lxa9867/imagefolder, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

run_coco Not yet run
unverified (harvested, no recorded run)
ivonajdenkoska/tulip/eval/coco.py
code served (permissive licence) · get_code("fef28a6f8d2d2a7a")
How it ran

no invocation recorded.

run_dci_long_openclip Not yet run
unverified (harvested, no recorded run)
ivonajdenkoska/tulip/eval/dci_long.py
code served (permissive licence) · get_code("4a7520e4ce7edf46")
How it ran

no invocation recorded.

run_sharegpt4v Not yet run
unverified (harvested, no recorded run)
ivonajdenkoska/tulip/eval/sharegpt4v.py
code served (permissive licence) · get_code("365d03d6a9cc44e6")
How it ran

no invocation recorded.

run_sharegpt4v_openclip Not yet run
unverified (harvested, no recorded run)
ivonajdenkoska/tulip/eval/sharegpt4v.py
code served (permissive licence) · get_code("388875d0199845b5")
How it ran

no invocation recorded.

Also present in other papers' code

Of 20 samples lifted from this paper's linked repositories, 20 carry a function identity and 0 do not. Of the 20 distinct functions with an identity, 10 are also present in other papers whose code we hold (10 with at least one paper dated earlier, 8 the same month, 8 later) and 10 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.

FunctionAlso present inEarliest-datedOthers named
get_cast_dtype 66 other papers
42 dated earlier, 6 the same month, 17 later, 1 same year, order not determinable
2103.00020 (2021-03)
Learning Transferable Visual Models From Natural Language Supervision
2210.08402 (2022-10)
LAION-5B: An open large-scale dataset for training next generation image-text models
2212.07143 (2022-12)
Reproducible scaling laws for contrastive language-image learning
and 63 more not named here
gather_features 57 other papers
40 dated earlier, 5 the same month, 11 later, 1 same year, order not determinable · 1 only as an assembly brick
2206.01670 (2022-06)
Egocentric Video-Language Pretraining
2212.00638 (2022-12)
Finetune like you pretrain: Improved finetuning of zero-shot vision models
2301.07094 (2023-01)
Learning Customized Visual Models with Retrieval-Augmented Knowledge
and 54 more not named here
convert_to_custom_text_state_dict 49 other papers
34 dated earlier, 4 the same month, 10 later, 1 same year, order not determinable
2103.00020 (2021-03)
Learning Transferable Visual Models From Natural Language Supervision
2210.08402 (2022-10)
LAION-5B: An open large-scale dataset for training next generation image-text models
2212.07143 (2022-12)
Reproducible scaling laws for contrastive language-image learning
and 46 more not named here
register_pooler 45 other papers
32 dated earlier, 3 the same month, 9 later, 1 same year, order not determinable
2301.07094 (2023-01)
Learning Customized Visual Models with Retrieval-Augmented Knowledge
2302.03668 (2023-02)
Hard Prompts Made Easy: Gradient-Based Discrete Optimization for Prompt Tuning and Discovery
2303.14814 (2023-03)
WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation
and 42 more not named here
prepare_inputs_for_generation 43 other papers
27 dated earlier, 3 the same month, 12 later, 1 same year, order not determinable
2301.07094 (2023-01)
Learning Customized Visual Models with Retrieval-Augmented Knowledge
2303.14814 (2023-03)
WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation
2305.07017 (2023-05)
An Inverse Scaling Law for CLIP Training
and 40 more not named here
get_input_dtype 29 other papers
17 dated earlier, 3 the same month, 9 later
2103.00020 (2021-03)
Learning Transferable Visual Models From Natural Language Supervision
2210.08402 (2022-10)
LAION-5B: An open large-scale dataset for training next generation image-text models
2212.07143 (2022-12)
Reproducible scaling laws for contrastive language-image learning
and 26 more not named here
neighbour_exchange 19 other papers
11 dated earlier, 3 the same month, 5 later
2303.15343 (2023-03)
Sigmoid Loss for Language Image Pre-Training
2403.15378 (2024-03)
Long-CLIP: Unlocking the Long-Text Capability of CLIP
2403.16831 (2024-03)
UrbanVLP: Multi-Granularity Vision-Language Pretraining for Urban Socioeconomic Indicator Prediction
and 16 more not named here
neighbour_exchange_bidir 17 other papers
9 dated earlier, 3 the same month, 5 later
2303.15343 (2023-03)
Sigmoid Loss for Language Image Pre-Training
2403.15378 (2024-03)
Long-CLIP: Unlocking the Long-Text Capability of CLIP
2403.16831 (2024-03)
UrbanVLP: Multi-Granularity Vision-Language Pretraining for Urban Socioeconomic Indicator Prediction
and 14 more not named here
natural_key 2 other papers
2 dated earlier
2306.09200 (2023-06)
ChessGPT: Bridging Policy Learning and Language Modeling
2407.11867 (2024-07)
Targeted Unlearning with Single Layer Unlearning Gradient
get_latest_checkpoint 1 other paper
1 dated earlier
2306.09200 (2023-06)
ChessGPT: Bridging Policy Learning and Language Modeling

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

We address the challenge of representing long captions in vision-language models, such as CLIP. By design these models are limited by fixed, absolute positional encodings, restricting inputs to a maximum of 77 tokens and hindering performance on tasks requiring longer descriptions. Although recent work has attempted to overcome this limit, their proposed approaches struggle to model token relationships over longer distances and simply extend to a fixed new token length. Instead, we propose a generalizable method, named TULIP, able to upgrade the token length to any length for CLIP-like models. We do so by improving the architecture with relative position encodings, followed by a training procedure that (i) distills the original CLIP text encoder into an encoder with relative position encodings and (ii) enhances the model for aligning longer captions with images. By effectively encoding captions longer than the default 77 tokens, our model outperforms baselines on cross-modal tasks such as retrieval and text-to-image generation. The code repository is available at https://github.com/ivonajdenkoska/tulip.

For agents

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

get_harvested_code_for_paper("2410.10034")
get_code_for_paper("2410.10034")
have("2410.10034")

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

curl https://syntology.ai/api/ran/2410.10034.json

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

[![Syntology run record](https://syntology.ai/api/ran/2410.10034.svg)](https://syntology.ai/paper/2410.10034)

Connect an agent — have() is free.