Joshyxwa/techjam2026

TechJam 2026

Dataset Summary

This public research draft contains 44,671 canonical images for binary AI-image authenticity experiments. Every row has a portable image path, fixed split, source and generator fields, lineage identifiers, integrity hashes, image properties, and processing history. It supports reproducible baselines and error analysis; it is not a deployment benchmark or a real-world accuracy claim.

Composition

Labels

Value	Images
ai_full	16,906
real	27,765

Splits

Value	Images
calibration	5,585
dev	6,091
own_locked	960
train	32,035

Source Datasets

Value	Images
amazon_berkeley_objects	3,000
blender_open_productions	1,536
dollar_street	117
local_generation	8,500
met_open_access	2,000
mirflickr_25k	6,000
nvidia_generation	8,406
open_food_facts	962
sid_set	13,999
wikimedia_commons	151

Files

File	Purpose
images/<split>/<label>/<asset_id>.<ext>	Canonical image files
labels.csv	Flat label and error-analysis table
manifest.parquet	Full machine-readable provenance and integrity manifest
metadata.jsonl	Compact loader-friendly index
verification.json	Package-level verification report

Labels And Splits

label is binary: real denotes a source-labelled non-synthetic image and ai_full denotes a source-labelled fully generated image. Preserve the supplied split assignments: use train to fit a model, dev for model selection and error analysis, and calibration only for threshold or probability calibration. Do not split lineage relatives or transform derivatives across these partitions.

Error Analysis

labels.csv is the recommended flat analysis interface. It combines the target, split, source/generator family, lineage and pair IDs, provenance, dimensions, format, hashes, and transform fields. Evaluate slices such as source_dataset, source_family, model_family, ai_subtype, file_format, resolution buckets, transform_chain, and watermark-review state alongside aggregate metrics.

import pandas as pd

labels = pd.read_csv("labels.csv")
counts = labels.groupby(["source_dataset", "label"]).size()

Reproducibility And Scope

manifest.parquet preserves the complete row schema and labels.csv exposes a stable, spreadsheet-friendly subset. The repository excludes raw source archives, credentials, checkpoints, and model caches. Upstream source and licence metadata is retained per image; users are responsible for following the recorded source terms. This draft has known source, format, and generator-family biases, so a random-split score should not be interpreted as robustness to unseen generators or distribution shift.