README.md

---
pretty_name: TechJam 2026 Data Draft
license: other
size_categories:
- 1K<n<10K
task_categories:
- image-classification
task_ids:
- multi-class-image-classification
tags:
- ai-generated-image-detection
- image-forensics
- research-only
- private-dataset
---

# TechJam 2026 Data Draft

## Dataset Summary

This **private research draft** validates a binary image-classifier pipeline:
the local data loader, feature extractor, classifier head, calibration step,
and robustness evaluation. It is not a public benchmark, a claim of real-world
detection quality, or part of the production 80k-master corpus.

The package has **10,000 canonical RGB PNG images**. Every image has a
portable manifest record with a label, split, source provenance, hashes, image
properties, and processing history.

## Composition

| Source | Real | Fully AI | Total | Notes |
| --- | ---: | ---: | ---: | --- |
| SID-Set | 2,500 | 2,500 | 5,000 | Canonical local snapshot; source-provided real/synthetic label |
| WildFake | 2,500 | 2,500 | 5,000 | Deterministic subset from complete local source archives |
| **Total** | **5,000** | **5,000** | **10,000** | Exact class and source balance |

WildFake real examples are evenly selected from `afhq`, `celebahq`, `church`,
`ffhq`, and `imagenet`. Fully-AI examples are evenly selected from `adm`,
`ddim`, `ddpm`, `imagen`, and `gan_based`. Each WildFake family contributes
500 images. COCO and DALL-E families are excluded.

## Splits

| Split | Real | Fully AI | Total | Intended use |
| --- | ---: | ---: | ---: | --- |
| `train` | 3,500 | 3,500 | 7,000 | Fit an architecture or classifier head |
| `dev` | 750 | 750 | 1,500 | Select settings and inspect errors |
| `calibration` | 750 | 750 | 1,500 | Set thresholds or fit probability calibration only |

Do not move records between splits or train on `dev`/`calibration` images or
their derivatives. The builder rejects cross-split collisions by asset ID,
base ID, lineage ID, exact SHA-256, exact perceptual hash, and near perceptual
hash distance at most five during WildFake selection.

## Labels

| Value | Meaning | Source evidence |
| --- | --- | --- |
| `real` | A source-labelled non-synthetic photograph/image | SID source path role or WildFake real archive role |
| `ai_full` | A source-labelled, fully generated AI image | SID source path role or WildFake generator archive role |

Labels are source-provided, not manually re-annotated. `label_evidence` and
`label_confidence` are present for every row; the confidence is `0.8` for this
draft. There are no `partial_ai`, `recaptured_ai`, or edited-image labels.

## Files And Metadata

| File | Purpose |
| --- | --- |
| `images/<split>/<label>/<asset_id>.png` | Canonical RGB image files |
| `manifest.parquet` | Full provenance, integrity, quality, and image-property manifest |
| `metadata.jsonl` | Compact image/label/split/source index for simple loaders |
| `verification.json` | Completed local decode, hash, and split-leakage audit |

`manifest.parquet` includes these field groups:

| Group | Key fields |
| --- | --- |
| Identity and split | `asset_id`, `base_id`, `lineage_id`, `split`, `label` |
| Label traceability | `label_evidence`, `label_confidence`, `real_subtype`, `ai_subtype` |
| Source provenance | `source_dataset`, `source_family`, `source_uri`, `source_archive`, `archive_member` |
| Image integrity | `sha256`, `phash`, `width`, `height`, `aspect_ratio`, `file_format`, `file_size_bytes` |
| Processing and quality | `processing_track`, `processing_history_status`, `provenance_status`, `quality_status` |
| Rights policy | `licence`, `licence_name`, `allowed_for_training`, `allowed_for_public_demo` |

Paths in the manifest are relative to this dataset repository. After download:

```python
from pathlib import Path

import pandas as pd
from PIL import Image

root = Path("path/to/data_draft")
manifest = pd.read_parquet(root / "manifest.parquet")
row = manifest.iloc[0]
image = Image.open(root / row.path).convert("RGB")
label = row.label  # "real" or "ai_full"
```

## Quality Checks

The packaged manifest passed the following local checks:

- 10,000 of 10,000 images decode as canonical RGB files.
- Recorded SHA-256, perceptual hash, width, and height match every image file.
- No detected cross-split collision by base ID, lineage ID, parent ID, exact
  hash, or exact perceptual hash.
- Final class counts are exactly balanced in every split.
- The package contains no raw source archives, API keys, model checkpoints,
  production generations, COCO records, or DALL-E records.

## Provenance And Restrictions

SID-Set rows retain upstream source URI and CC-BY-4.0 metadata. WildFake rows
retain the exact ZIP archive and member path but are marked
`licence_audit_required`. Therefore:

- Keep this repository private.
- Do not redistribute WildFake-derived files or make public demos from them
  until the upstream licence review is complete.
- Do not add these records to the production 80k corpus or use them for
  benchmark/generalization claims.

The `other` licence tag describes the mixed, restricted state of the package;
it grants no rights beyond the underlying source datasets.

## Limitations

This small mixture is useful for pipeline and architecture smoke tests, but it
is source-family constrained. Its results may be optimistic because training,
development, and calibration all originate from the same two upstream datasets.
It does not model social-media reposting, video-frame extraction, partial edits,
recapture, contemporary held-out generators, or real-world class prevalence.
Report per-source and per-transform errors, not only aggregate AUROC/AUPRC.

## Reproducibility

The deterministic selection seed is `20260830`. `manifest.parquet` preserves
the selection and processing lineage, stable asset/base/lineage identifiers,
and content hashes required to audit a local rebuild before training.