blanchon/cs2_dataset_render_part2

OpenCS2 — POV Renders

Browse with the OpenCS2 Viewer — every match, map and round, with all 10 player POVs synced on one timeline.

Tick-aligned Counter-Strike 2 POV training clips, rendered from blanchon/cs2_dataset_demo. Each row is ≤1 minute of one player's perspective; ten POVs per round share the same tick clock.

Per chunk:

• Video — 1280×720 @ 32 fps, near-lossless H.264.
• Audio — per-player stereo, mixed from that player's position and orientation.
• Inputs — every tick: keys, mouse delta, view angles, fire/jump/use, weapon switches.
• World state — every tick, all 10 players: position, velocity, view, health, armor, weapon, alive flag.

---

Usage

Four configs. previews is the cheap default; chunks is the heavy training data.

Config	Row	Use
previews (default)	low-res preview.mp4 + 1 Hz inputs/world sidecars	browsing, sanity checks
chunks	path-only video.mp4 + audio.wav, embedded inputs/worlds	training
matches	one per (match_id, map_name) with team/event metadata	filtering / index
rounds	one per (match_id, map_name, round) with tick boundaries	filtering / index

Stream with datasets

from datasets import load_dataset

# Default browsing view
previews = load_dataset("blanchon/cs2_dataset_render_part2", split="train", streaming=True)

# Training rows. Pass columns/filters at load time so the parquet reader pushes them down.
chunks = load_dataset(
    "blanchon/cs2_dataset_render_part2", "chunks",
    split="train", streaming=True,
    columns=["video", "audio", "inputs", "worlds", "match_id", "round", "player"],
    filters=[("player", "==", 0)],
)

Query with DuckDB

INSTALL httpfs; LOAD httpfs;

-- Match index
SELECT match_id, map_name, team1, team2, event, match_date
FROM 'hf://datasets/blanchon/cs2_dataset_render_part2/index/manifest-*.parquet'
WHERE event ILIKE '%IEM%';

-- Long rounds only
SELECT match_id, map_name, round, round_duration_ticks
FROM 'hf://datasets/blanchon/cs2_dataset_render_part2/index/rounds-*.parquet'
WHERE round_duration_ticks > 3000;

-- Preview rows, filter by AWP plays (cheap visual triage)
SELECT match_id, round, player, chunk_index, primary_weapon
FROM 'hf://datasets/blanchon/cs2_dataset_render_part2/data/**/chunks-preview-*.parquet'
WHERE primary_weapon = 'AWP';

Partial download with the hf CLI

Hive partitioning lets you target a single match/map/player without touching the rest:

# Index files only (~MB, scan-friendly)
hf download blanchon/cs2_dataset_render_part2 --repo-type dataset --include "index/*.parquet"

# All previews for one match
hf download blanchon/cs2_dataset_render_part2 --repo-type dataset \
    --include "data/match_id=2393343/**/chunks-preview-*.parquet" \
    --include "data/match_id=2393343/**/previews/**"

# Full chunks (video+audio+parquet) for one player on one map
hf download blanchon/cs2_dataset_render_part2 --repo-type dataset \
    --include "data/match_id=2393343/map_name=de_ancient/player=0/**"

---

Structure

Repository layout

data/
  match_id=<id>/map_name=<map>/player=<0-9>/
    chunks-preview-<machine>-<uuid>.parquet
    chunks-full-<machine>-<uuid>.parquet
    chunks/chunk_<n>/{video.mp4, audio.wav}
    previews/chunk_<n>/{preview.mp4, inputs.preview.json, world.preview.jsonl}
index/
  manifest-<machine>-<uuid>.parquet     # one row per (match, map)
  rounds-<machine>-<uuid>.parquet       # one row per (match, map, round)

Hive key=value directories prune at the path level. Media files are loose; the parquets store relative *_path columns plus an hf://... URI for the Hub dataset viewer. Parallel render workers each write their own <machine>-<uuid> shard so uploads never collide.

Row semantics

• player is the canonical 0-9 player index, stable for a match.
• spec_slot is the transient CS2 spectator slot used to capture the POV — useful only for debugging.
• Recording starts at the playable round start (freeze_end_tick). Death POVs keep a short death tail clamped to round end; survivor POVs keep a short post-round tail, so final kill/round-win feedback is not clipped. POVs in the same round can have different durations.
• inputs and worlds are struct-of-arrays in the chunks parquet — row["worlds"]["X"] returns the per-tick X position list for all 10 players.
• Hot filter columns on chunks: match_id, map_name, player, round, chunk_index, primary_weapon, player_side, survived_chunk, damage_taken, shots_fired, distance_traveled, weapons_used.

---

Creation

Built with the recorder in https://github.com/julien-blanchon/opencs2-dataset:

1. Demos — pulled from blanchon/cs2_dataset_demo.
2. Render — a headless CS2 + custom plugin replays each demo, captures every player POV tick-by-tick, and streams raw frames to NVENC. Validated end-to-end: start_tick == requested, frame count matches the tick range.
3. Parquet — chunks (≤1 min) sorted by (round, chunk_index), written with row_group_size=1, write_page_index=True and use_content_defined_chunking=True so re-uploads of touched chunks transfer near-zero bytes via Xet.
4. Upload — each render worker writes its own <machine>-<uuid> shard via upload_large_folder with hf_transfer + hf_xet.

---

Licensing & Usage

.dem source data is mirrored from HLTV; downstream use is bound by the original tournament terms. Renders + metadata: CC-BY-4.0.

---

Citation

@misc{blanchon2026opencs2,
  author       = {Julien Blanchon},
  title        = {OpenCS2 Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://github.com/julien-blanchon/opencs2-dataset}}
}