sneakyzero/cs2-pro-round-economy
CS2 Professional Round Economy A compact, researcher-oriented extraction of round-level economy and freeze-time inventory from 1,988 professional Counter-Strike 2 maps. The source demos total 935.2 GB; this release contains only approximately 22.2 MiB of compressed Parquet tables and does not redistribute any demo files. The maps come from 860 series across 70 events collected in blanchon/cs2_dataset_demo, fixed at source revision 912a363098feec3f220e41d68634d2f71676c28b. They… See the full description on the dataset page: https://huggingface.co/datasets/sneakyzero/cs2-pro-round-economy.
CS2 Professional Round Economy
A compact, researcher-oriented extraction of round-level economy and freeze-time inventory from 1,988 professional Counter-Strike 2 maps. The source demos total 935.2 GB; this release contains only approximately 22.2 MiB of compressed Parquet tables and does not redistribute any demo files.
The maps come from 860 series across 70 events collected in `blanchon/cs2_dataset_demo`, fixed at source revision `912a363098feec3f220e41d68634d2f71676c28b`. They were played from 2026-03-23 through 2026-04-20 on patch 14141 and were originally indexed from professional matches listed by HLTV.
Why this dataset exists
Raw CS2 demos are expensive to download and parse when the research question only needs economy and inventory. This dataset exposes the reusable factual layer: cash, spending, equipment value, armor, helmets, defuse kits, item inventories, round outcomes, map metadata, and provenance.
The base tables deliberately do not label rounds as eco, half buy, or force buy. Those terms require researcher-selected thresholds and can change with the question being studied. Continuous purchase and equipment fields are retained so users can define and version their own policies.
Contents
Primary keys, field definitions, and interpretation notes are documented in `DATA_DICTIONARY.md`. Extraction coverage and quality-control results are summarized in `EXTRACTION_REPORT.md`, with machine-readable checks in `audit/full_audit.json`.
Loading the data
Each table has a different schema and is therefore exposed as a separate dataset configuration.
from datasets import load_dataset
economy = load_dataset(
"sneakyzero/cs2-pro-round-economy",
"player_round_economy",
split="train",
)
opportunities = load_dataset(
"sneakyzero/cs2-pro-round-economy",
"pistol_loss_opportunities",
split="train",
)Parquet-aware tools can also read files directly:
import duckdb
rounds = duckdb.sql("""
SELECT *
FROM 'hf://datasets/sneakyzero/cs2-pro-round-economy/data/full_rounds.parquet'
""").df()Extraction and provenance
- Upstream repository:
blanchon/cs2_dataset_demo - Pinned upstream revision:
912a363098feec3f220e41d68634d2f71676c28b - Demo parser:
demoparser2==0.42.0 - Export schema version:
0.1.0 - Snapshot timing: freeze time end for each round
- Raw demos retained or redistributed: no
Stable map keys, upstream file paths, source revision, parser version, and demo SHA-256 hashes are preserved where applicable. This supports traceability without hosting the original demos.
Quality control
All 1,988 maps pass structural QC. The release has no duplicate declared primary keys, all round counts match the pinned manifest, and all 3,960 decision snapshots are complete. The automated release audit passed 29/29 assertions.
Two maps contain one genuine player snapshot each where no active pawn existed at freeze end. Those player rows are retained with player_active_at_freeze_end=false, and the maps are marked qc_status=review. Eleven maps are conservatively excluded from the original study's versioned analysis population; they remain in the base data with explicit analysis_eligible_v1 and exclusion-reason fields.
Scope and limitations
- This is professional-match data from the upstream collection, not matchmaking or FACEIT public-queue data.
- Coverage is a one-month snapshot of one game patch, not a complete history of professional CS2.
- Team names occasionally cannot be resolved safely from demo clan tags. Unresolved values are left missing rather than guessed.
- Public professional-player Steam IDs and in-demo names are included because they are stable join keys present in publicly distributed match demos. The dataset does not attempt to infer private identity information.
pistol_loss_opportunitiesincludes reproducible derived features and short-horizon outcomes, but no final strategy labels.
License and source terms
This release is provided under CC BY 4.0, matching the license metadata on the upstream dataset. Please attribute both this derived dataset and `blanchon/cs2_dataset_demo`.
The upstream card states that its demos are mirrored from HLTV's public demo-download endpoint and were provided by tournament organizers. Users remain responsible for any terms applicable to the original tournament demos and sites. This release distributes structured observations only, not demos, video, audio, logos, or other original media.
中文简介
这是从1,988张职业CS2比赛地图中提取的逐回合经济与库存数据。发布包只包含压缩Parquet表,不包含约935 GB的原始demo。基础表保留连续现金、支出、装备和库存事实,不预先把购买策略标为eco、half buy或force,便于研究者自行定义并复现分析口径。
