ybashir/gamed-game-classification-dataset

Gamed — Game Classification Dataset

Train and held-out eval frames for a 3-class game classifier (cs2, dota2, valorant). Used to train and evaluate ybashir/gamed-game-classification, the §1.1 hot-path detector in coach-api.

The source .mp4 files are not included — see urls.txt for the original YouTube links. All frames here are derived at 0.4 fps via ffmpeg and resized to 224×224 (squashed 1:1).

Structure

train/
  manifest.csv        6450 rows — path,class,video_id,frame_idx,t_sec
  frames_224.zip      6450 JPGs (224×224), grouped by class inside the zip

eval/
  manifest.csv        450 rows  — 150/class evenly-spaced subsample
  manifest_full.csv   1118 rows — full extraction (cs2 390 / dota2 295 / valorant 433)
  frames/
    cs2/      390 JPGs
    dota2/    295 JPGs
    valorant/ 433 JPGs
  predictions.csv      Held-out predictions from the published checkpoint (450)
  predictions_full.csv Held-out predictions from the published checkpoint (1118)
  report.md            Acc + P/R/F1 + confusion matrix (450)
  report_full.md       Acc + P/R/F1 + confusion matrix (1118)

urls.txt              YouTube source URLs (train + eval) + per-video start offsets

Manifest schema

Both train/manifest.csv and eval/manifest.csv have the same columns:

path,class,video_id,frame_idx,t_sec[,split]

• path — relative to the original repo cwd (gamed/)
• class — cs2 / dota2 / valorant
• video_id — YouTube ID (11 chars), parsed from the source filename prefix
• frame_idx — 1-indexed JPG index within the video
• t_sec — timestamp in the source video the frame was sampled from
• split — only on eval manifests, always eval

Train vs eval — important caveat

Training was done with a frame-level random_split(0.70/0.15/0.15) on train/manifest.csv. The same video_ids appear in train and val/test, so the 99% accuracy quoted in the original notebook is over-optimistic (video-level leakage).

The eval/ split fixes this:

• 3 fresh YouTube videos (different video_id, different uploaders)
• Zero video_id overlap with the training set
• Reported eval accuracy: 99.91 % on 1118 frames, 100 % on the 150/class subsample

This is the number to cite. Tool to reproduce: cli/gamed_classification_eval.py.

How frames were extracted

scripts/extract_frames.py in the repo runs ffmpeg at 0.4 fps with optional per-video --trim-start offsets to skip menus / intros. Frames are JPEG q=5 (~85% quality) at 224×224 (squashed).

Per-video offsets used for the eval set:

Class	Video ID	start_sec
cs2	FF1pN66xVfA	0
dota2	2fpY4OeoY0c	210
valorant	Alrz2LCm8pg	0

Eval results from the published checkpoint

See eval/report.md and eval/report_full.md for the full numbers. Headline:

Set	Frames	Accuracy
eval/manifest.csv (450)	150 / class	100 %
eval/manifest_full.csv (1118)	390 / 295 / 433	99.91 %

Single error: valorant/Alrz2LCm8pg_000133.jpg (t≈330 s) — a round-end "WON" overlay misread as CS2 because Valorant's orange/red Combat Report panel mimics CS2's post-round MVP card.

License

Frames are 224×224 derivatives of publicly-uploaded "no commentary" YouTube gameplay videos. CC-BY-4.0 on the derived dataset itself; the underlying game footage remains the property of the original uploaders and the respective game publishers (Valve, Riot Games).

Citation

Repo: https://github.com/bashiryounis/gamed