The Dataset Viewer has been disabled on this dataset.

TikTok TechJam 2026 — train sources

How to reproduce the Seer hero mix (configs/seer_vitl_512.yaml). Most sources already live on the Hub; only the filtered slices were re-uploaded.

Mixture source Class Get it from Notes
comfor mixed OwensLab/CommunityForensics-Small full 186-shard train dump
ntire mixed deepfakesMSU/NTIRE-RobustAIGenDetection-train all 6 shards
openfake mixed glennwuwu/tiktok-techjam-2026-openfake-train ranked 30-gen + Pexels/LAION slice, not the 3.44 TB repo
flux-reason fake LucasFang/FLUX-Reason-6M local cache was shards 0000000063 of Aesthetics-Part01
frontier-fakes fake julienlucas/midjourney-dalle-sd-nanobananapro-dataset invert labels, keep_label=1
sid-set fake saberzl/SID_Set keep class 1 (full synthetic) only
gs-images-v3 fake gasstation/gs-images-v3 then scripts/wire_gasstation.py --versions v3
gs-images-v4 fake gasstation/gs-images-v4 then scripts/wire_gasstation.py --versions v4
laion400m-1 real glennwuwu/tiktok-techjam-2026-laion400m-reals min side >512, first 17 shards, ~400k JPEGs
open-images-v7 real Open Images V7 val+test scripts/download_open_images.py

Eval set (not train): glennwuwu/tiktok-techjam-2026-eval (COCO val2017 + WildFake DALL·E Advanced).

Downloads last month
33