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. The default config is previews, a small path-only overlay+audio video view for browsing. The heavy chunks config contains full-resolution video/audio as loose files referenced by path, plus embedded inputs and world-state streams.

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

Configs

• previews (default): one low-resolution preview_video row per chunk with the input overlay baked in. Preview MP4s are stored as loose files and the Parquet stores only their relative paths, so filtering stays cheap. The preview_path column is relative to the preview Parquet directory (previews/chunk_000123/preview.mp4), while preview_video.path is an hf://datasets/<repo>@main/... URI so Hugging Face's dataset viewer can resolve it without inlining bytes.
• matches: one row per rendered (match_id, map_name) with team/event/date metadata.
• rounds: one row per rendered (match_id, map_name, round) with tick boundaries.
• chunks: full training rows with path-only video/audio, embedded inputs/worlds, and typed metadata.

Filesystem Layout

data/
  match_id=<match_id>/map_name=<map_name>/player=<player>/
    chunks-preview-<machine>-<uuid>.parquet
    chunks-full-<machine>-<uuid>.parquet
    chunks/
      chunk_<ordinal>/
        video.mp4
        audio.wav
    previews/
      chunk_<ordinal>/preview.mp4
index/
  manifest-<machine>-<uuid>.parquet
  rounds-<machine>-<uuid>.parquet
state/
  processed/<input_metadata_stem>/<match_id>.json
  failed/...
  skipped/...

Every machine writes a unique <machine>-<uuid> shard, so parallel uploads do not touch the same files. state/processed is written only after the shard upload succeeds. Data files use Hive-style key=value directories for pruning. The same keys are also stored inside the Parquets so the Hugging Face viewer and datasets.load_dataset("parquet", ...) expose them even without applying Hive partitioning. Parquets use best-effort bloom filters on hot filter columns.

Stream With datasets

from datasets import load_dataset

repo = "blanchon/cs2_dataset_render"

# Cheap default browsing view.
previews = load_dataset(repo, split="train", streaming=True)
for row in previews.take(3):
    print(row["match_id"], row["round"], row["player"], row["preview_video"])

# Full training rows. Use columns/filters to avoid pulling unused bytes.
chunks = load_dataset(
    repo,
    "chunks",
    split="train",
    streaming=True,
    columns=["video", "audio", "inputs", "worlds", "match_id", "round", "player"],
    filters=[("player", "==", 0)],
)

Query With DuckDB

-- Match-level index only; no media bytes.
SELECT match_id, map_name, team1, team2, event, match_date
FROM 'hf://datasets/blanchon/cs2_dataset_render/index/manifest-*.parquet'
LIMIT 20;

-- Round timing index only.
SELECT match_id, map_name, round, round_duration_ticks
FROM 'hf://datasets/blanchon/cs2_dataset_render/index/rounds-*.parquet'
WHERE round_duration_ticks > 3000;

-- Preview rows for fast visual review.
SELECT match_id, map_name, round, player, chunk_index, primary_weapon
FROM 'hf://datasets/blanchon/cs2_dataset_render/data/**/chunks-preview-*.parquet'
WHERE player = 0
LIMIT 20;

Partial Download

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

hf download blanchon/cs2_dataset_render --repo-type dataset \
  --include "data/match_id=2393343/**/chunks-preview-*.parquet"

hf download blanchon/cs2_dataset_render --repo-type dataset \
  --include "data/match_id=2393343/map_name=de_ancient/player=0/chunks-full-*.parquet"

Row Semantics

• player is the stable canonical 0-9 player index for a match.
• spec_slot is the transient CS2 spectator slot used only to record the POV.
• Recording starts at the playable round start (freeze_end_tick) and stops at the player's death tick, or at round end for survivors.
• The production worker validates exactly 10 canonical players per round, unique spec-slot resolution, no recording past death, and valid video/audio/ inputs/world sidecars before upload.