blanchon/cs2_dataset_render

HLTV CS2 POV Rendered Dataset

Rendered Counter-Strike 2 POV training clips derived from blanchon/cs2_dataset_demo. Each row is one one-minute-or-shorter player POV chunk with loose files referenced from metadata:

• file_name: video.mp4
• audio_file_name: tick-aligned audio.wav
• inputs_file_name: per-tick input/action inputs.parquet
• worlds_file_name: per-tick world-state worlds.parquet
• metadata_file_name: chunk-local metadata.json
• artifact_file_names: list of all per-chunk artifacts for Hub preview / partial download tooling

The raw HLTV .dem files stay in the source dataset. This repo stores the rendered training artifacts only.

Layout

metadata-<machine>-<uuid>.parquet
clips/
  shard-<machine>-<uuid>/
    <match_id>/<map>/round_00015/player_00003/chunk_000000/
      video.mp4
      audio.wav
      inputs.parquet
      worlds.parquet
      metadata.json
source/
  shard-<machine>-<uuid>/<match_id>/
    meta.json
    <demo>.analysis.json
state/
  processed/<input_metadata_stem>/<match_id>.json

Rows are safe to append from many machines because every upload uses a unique machine/UUID shard.

Query metadata without downloading clips

FROM 'hf://datasets/blanchon/cs2_dataset_render/metadata-*.parquet'
SELECT match_id, map_name, round, player, duration_s, file_name
LIMIT 10;

Stream with datasets

from datasets import load_dataset, Video, Audio

ds = load_dataset("blanchon/cs2_dataset_render", split="train", streaming=True)
ds = ds.cast_column("file_name", Video(decode=False))
ds = ds.cast_column("audio_file_name", Audio(decode=False))

for row in ds.take(1):
    print(row["match_id"], row["round"], row["player"], row["file_name"])

Partial download

hf download blanchon/cs2_dataset_render --repo-type dataset \
  --include "metadata-*.parquet"

hf download blanchon/cs2_dataset_render --repo-type dataset \
  --include "clips/shard-*/2393343/**"