feifeili00/rendertext_semleak

TSL60 Requests, Images, and Semantic-Leakage Labels

TSL60 contains 360 text-to-image requests. Each request has one generated image from each of six models: flux_2_dev, qwen_image, z_image_turbo, longcat, gpt_image_2, and nano_banana_2. TSL21 is a 126-request subset of TSL60.

Test semantic leakage with TSL60 requests

Download the benchmark CSVs without images:

hf download feifeili00/rendertext_semleak --repo-type dataset --include "data/**" --local-dir .

Use the prompt column in data/tsl60/TSL60_requests.csv to generate images with a target model. The id column identifies each request. The other CSVs provide the following inputs and reference information:

CSV	Contents
data/tsl60/TSL60_requests.csv	Final generation prompts, request IDs, target goals, and rendered texts
data/tsl60/TSL60_goals.csv	Semantic-leakage goals and dimensions
data/tsl60/rendered_texts.csv	Candidate text associated with each goal
data/tsl60/TSL60_base_prompts.csv	Two base-image prompts for each main prompt
data/cvtg_templates.csv, data/synthetic_templates.csv, data/textinvision_templates.csv	Main prompt templates from the three source collections
data/tsl60/TSL60_word_of_interest.csv	Target word for each request
data/tsl21/TSL21_requests.csv, data/tsl21/TSL21_goals.csv	Requests and goals in the TSL21 subset
data/benign/cvtg100.csv, data/benign/tiv100.csv	Two 100-request benign sets with expected rendered text

To load a request table directly from the Hub:

from datasets import load_dataset

requests = load_dataset("feifeili00/rendertext_semleak", "requests", split="tsl60")

Validate semantic-leakage scorers against labeled images

Load the 2,160 images and their human and automatic labels:

from datasets import load_dataset

repo = "feifeili00/rendertext_semleak"
images = load_dataset(repo, "images", split="tsl60")
labels = load_dataset(repo, "labels", split="tsl60")
gsl = load_dataset(repo, "gsl", split="tsl60")

Join images and results by id and model. labels.csv contains separate three-class labels from anonymized annotators A and B: 0 = no leakage, 1 = weak or partial leakage, and 2 = clear semantic leakage. For images with both labels, the overall human ground truth is

Both annotators labeled all TSL60 images from the four open models and the TSL21 images from the two commercial models; unlabeled cells are empty and have no human ground-truth label.

gsl.csv covers every image. GSL asks two slightly different yes/no questions about whether the target goal appears in the non-text image content. Each question records its text and a structured response: answer (yes/no), confidence, score, and rationale. The overall result is

To evaluate on TSL21, select its request IDs from the same image and label tables:

ids21 = set(load_dataset(repo, "requests", split="tsl21")["id"])
images21 = images.filter(lambda row: row["id"] in ids21)
labels21 = labels.filter(lambda row: row["id"] in ids21)
gsl21 = gsl.filter(lambda row: row["id"] in ids21)

To download only the 756 TSL21 image files instead of all TSL60 images:

from datasets import load_dataset
from huggingface_hub import snapshot_download

repo = "feifeili00/rendertext_semleak"
ids21 = load_dataset(repo, "requests", split="tsl21")["id"]
snapshot_download(
    repo_id=repo, repo_type="dataset", local_dir=".",
    allow_patterns=[
        "data/tsl21/TSL21_requests.csv", "metadata.jsonl", "labels.csv",
        "gsl.csv", "read_dataset.py",
        *[f"images/*/{request_id}.png" for request_id in ids21],
    ],
)

For a local copy, python read_dataset.py tsl60 or python read_dataset.py tsl21 reports image counts and label coverage by model.