README.md

---
license: cc-by-4.0
task_categories:
  - tabular-classification
  - tabular-regression
  - feature-extraction
  - other
language:
  - en
tags:
  - cs2
  - counter-strike
  - esports
  - sports-analytics
  - competitive-gaming
  - machine-learning
  - elo
  - feature-engineering
  - win-prediction
  - sports-betting
  - tabular-data
size_categories:
  - 10K<n<100K
pretty_name: CounterQuant CS2 GoldLite
---

# CounterQuant CS2 GoldLite

**ML-ready feature tables for professional CS2 match outcome modelling — Elo ratings, team form, head-to-head records, and event context. No black boxes, no proprietary signals.**

> ⭐ If you use this dataset in research, a product, or any publication,
> please **cite the author** (see [Citation](#citation) below).

GoldLite is the open, reproducible subset of CounterQuant's internal feature pipeline. It provides pre-computed, match-level features derived entirely from public match results and event metadata — no demo parsing required, no market data, no model outputs. Train your own win-prediction model on a decade of professional CS2.

Curated and maintained by **[Eimantas Kulbe (KEDevO)](https://counterquant.com)** as part of the **CounterQuant** esports intelligence platform.

---

## The CounterQuant Data Stack

GoldLite is the top public-facing tier of a four-layer architecture:

| Tier | Dataset | Contents | Status |
|------|---------|----------|--------|
| **Raw demos** | [CounterQuant CS2 Demos](https://huggingface.co/datasets/KEDevO/CounterQuant-CS2-Demos) | Raw `.dem` files | Live |
| **Bronze** | [CounterQuant CS2 Bronze](https://huggingface.co/datasets/KEDevO/CounterQuant-CS2-Bronze) | Tick-level events: kills, damages, flashes, utility | Live, growing |
| **Silver** | [CounterQuant CS2 Silver](https://huggingface.co/datasets/KEDevO/CounterQuant-CS2-Silver) | Match results, map scores, player stats, rosters | Staging for export |
| **Gold Lite** ← *this dataset* | — | Elo, form, H2H, event context — ML features | Staging for export |

The full internal Gold layer (demo-derived features, market signals, 509-feature ML vectors) is proprietary and powers the CounterQuant prediction API. GoldLite is the public, reproducible subset of that pipeline.

---

## Dataset at a Glance

| Metric | Value |
|--------|-------|
| **Team Elo records in DB** | 45,502 (per team, per map pool) |
| **Matches with features** | Growing — 2024 first public batch |
| **Year range (planned)** | 2012 – 2024 (first release) |
| **Feature count (GoldLite)** | ~50 pre-match features |
| **Full pipeline feature count** | 509 (internal, not released) |
| **Format** | Parquet (Snappy compression) |
| **License** | CC BY 4.0 |

---

## Current State & Release Schedule

> **GoldLite is not yet published to HuggingFace as Parquet files.**
> Elo ratings are computed and live in CounterQuant's database.
> The pre-match feature export is in progress, with 2024 as the first public batch.

| Release | Coverage | Target Date | Notes |
|---------|----------|-------------|-------|
| **v1 — 2024** | Jan 2024 – Dec 2024 | Q3 2026 | First public release, alongside Silver |
| **Backfill — 2012–2023** | 2012 – 2023 | Q4 2026 | Historical Elo + features |
| **v2 — 2025** | 2025 | 2027 | 2-year delay policy |
| **2026+** | 2026+ | 2028+ | 2-year delay minimum |

---

## What's in GoldLite

GoldLite contains **pre-match features computed from historical match results** — the information that would be available to a predictor at kick-off. All features are constructed from public sources only.

### Feature categories

| Category | Features | Source |
|----------|----------|--------|
| **Elo ratings** | Global team Elo, per-map Elo, Elo uncertainty, Elo differential | Computed from Silver match results |
| **Recent form** | Win rate last 5/10/20 matches, map differential last 10, days since last match | Silver matches |
| **Head-to-head** | All-time H2H record, last 12 months H2H, H2H on specific map | Silver matches |
| **Map pool** | Team win rate per map (last 30 games), map familiarity score, most/least played maps | Silver maps |
| **Event context** | Prize pool tier, LAN vs online, event stage (groups/playoffs/final), days to event end | Silver events |
| **Roster stability** | Avg days together, roster change flag (last 30 days) | Silver rosters |
| **Player aggregate** | Avg team rating last 20 maps, avg ADR last 20 maps, avg KAST last 20 maps | Silver player_stats |

### What GoldLite does NOT include

| Excluded | Where it lives |
|----------|---------------|
| Demo-derived features (eco win rates, KAST, trade kills, utility efficiency) | Internal Gold (CounterQuant API) |
| Polymarket / betting market odds | Private, never released |
| XGBoost/LightGBM model output probabilities | Private (CounterQuant predictions) |
| 509-feature full ML vectors | Private (CounterQuant API) |
| Live round-by-round probability updates | Private (CounterQuant live feed) |
| Proprietary tactical embeddings | Private, never released |

If you need demo-derived features, use [Bronze](https://huggingface.co/datasets/KEDevO/CounterQuant-CS2-Bronze) and compute them yourself — that's exactly what CounterQuant does internally.

---

## Planned File Structure

```
data/
├── team_elo/
│   ├── team_elo_2024.parquet           # Elo snapshot at each match played in 2024
│   ├── team_elo_2012_2023.parquet      # Historical backfill
│   └── team_elo_current.parquet        # Latest Elo per team per map (updated)
├── match_features/
│   ├── match_features_2024.parquet     # Pre-match feature vector per 2024 match
│   └── match_features_2012_2023.parquet
├── player_rolling_stats/
│   ├── player_rolling_2024.parquet     # Rolling 5/10/20-match averages per player
│   └── ...
└── map_pool/
    ├── map_pool_stats_2024.parquet     # Team win rates per map, as of each match
    └── ...
```

---

## Schema Reference

### `team_elo_YYYY.parquet`

Elo rating for each team at the point of each match. One row per team per match.

| Column | Type | Description |
|--------|------|-------------|
| `match_id` | int64 | Match ID |
| `team_id` | int64 | Team ID |
| `team_name` | string | Team name (denormalized) |
| `map_name` | string | `all` for global Elo; CS2 map name for map-specific Elo |
| `elo_before` | float32 | Elo rating entering this match |
| `elo_after` | float32 | Elo rating after this match |
| `elo_change` | float32 | Delta (positive = won, negative = lost) |
| `uncertainty` | float32 | Elo uncertainty / confidence interval |
| `matches_played` | int32 | Total matches used to compute this rating |
| `date` | date | Match date |

---

### `match_features_YYYY.parquet`

Pre-match feature snapshot. One row per match, all features computed from data available **before** kick-off. Safe for train/test split — no data leakage.

| Column | Type | Description |
|--------|------|-------------|
| `match_id` | int64 | HLTV match ID |
| `date` | date | Match date |
| `tier` | int16 | Match tier (1/2/3) |
| `is_lan` | bool | LAN event |
| `prize_pool_usd` | int32 | Event prize pool in USD |
| `team1_id` | int64 | Team 1 ID |
| `team2_id` | int64 | Team 2 ID |
| `team1_elo` | float32 | Global Elo entering this match |
| `team2_elo` | float32 | Global Elo entering this match |
| `elo_diff` | float32 | team1_elo − team2_elo |
| `team1_form_5` | float32 | Win rate last 5 matches |
| `team2_form_5` | float32 | Win rate last 5 matches |
| `team1_form_10` | float32 | Win rate last 10 matches |
| `team2_form_10` | float32 | Win rate last 10 matches |
| `team1_map_diff_10` | float32 | Map differential (maps won − lost) last 10 |
| `team2_map_diff_10` | float32 | Map differential last 10 |
| `h2h_team1_wins` | int16 | All-time H2H wins for team 1 |
| `h2h_team2_wins` | int16 | All-time H2H wins for team 2 |
| `h2h_team1_wins_12m` | int16 | H2H wins last 12 months |
| `h2h_team2_wins_12m` | int16 | H2H wins last 12 months |
| `team1_days_since_match` | int16 | Days since team 1's last match |
| `team2_days_since_match` | int16 | Days since team 2's last match |
| `team1_avg_rating_20` | float32 | Avg team player rating last 20 maps |
| `team2_avg_rating_20` | float32 | Avg team player rating last 20 maps |
| `team1_avg_adr_20` | float32 | Avg team ADR last 20 maps |
| `team2_avg_adr_20` | float32 | Avg team ADR last 20 maps |
| `roster_change_team1_30d` | bool | Team 1 had a roster change in last 30 days |
| `roster_change_team2_30d` | bool | Team 2 had a roster change in last 30 days |
| `team1_won` | bool | **Label** — did team 1 win? (null = match not completed) |

---

### `player_rolling_YYYY.parquet`

Rolling player performance windows, computed at each match. Useful for player-level predictions and ranking models.

| Column | Type | Description |
|--------|------|-------------|
| `match_id` | int64 | Match ID |
| `player_id` | int64 | Player ID |
| `player_ign` | string | IGN |
| `team_id` | int64 | Team ID at this match |
| `maps_played_total` | int16 | Career maps played at this point |
| `rolling_kills_5` | float32 | Avg kills per map last 5 maps |
| `rolling_kills_10` | float32 | Avg kills per map last 10 maps |
| `rolling_adr_10` | float32 | Avg ADR last 10 maps |
| `rolling_adr_20` | float32 | Avg ADR last 20 maps |
| `rolling_kast_10` | float32 | Avg KAST last 10 maps |
| `rolling_rating_20` | float32 | Avg HLTV Rating 2.0 last 20 maps |
| `rating_trend_10` | float32 | Linear slope of rating over last 10 maps (form direction) |
| `days_since_last_match` | int16 | Days since player's last match |

---

### `map_pool_stats_YYYY.parquet`

Team win rates per map, computed at each match date (so you can join to match_features for map-specific models).

| Column | Type | Description |
|--------|------|-------------|
| `as_of_match_id` | int64 | The match this snapshot was computed for |
| `as_of_date` | date | Date of the snapshot |
| `team_id` | int64 | Team ID |
| `map_name` | string | CS2 map name |
| `maps_played` | int16 | Maps played on this map (last 2 years) |
| `maps_won` | int16 | Maps won |
| `win_rate` | float32 | Win rate (0–1) |
| `maps_played_tier1` | int16 | Maps played at Tier 1 only |
| `win_rate_tier1` | float32 | Win rate at Tier 1 |

---

## Quick Start (Preview)

Once 2024 data is published:

```python
import polars as pl
from huggingface_hub import hf_hub_download

# Download pre-match features for 2024
path = hf_hub_download(
    repo_id="KEDevO/CounterQuant-CS2-GoldLite",
    repo_type="dataset",
    filename="data/match_features/match_features_2024.parquet",
)
df = pl.read_parquet(path)

# Drop rows with incomplete data (ongoing / no-result matches)
df = df.filter(pl.col("team1_won").is_not_null())

X = df.drop(["match_id", "date", "team1_id", "team2_id", "team1_won"])
y = df["team1_won"].cast(pl.Int8)
print(f"Training samples: {len(X)}, Features: {X.width}")
```

```python
from sklearn.model_selection import train_test_split
from xgboost import XGBClassifier
from sklearn.metrics import brier_score_loss, roc_auc_score

X_train, X_test, y_train, y_test = train_test_split(
    X.to_pandas(), y.to_pandas(),
    test_size=0.2, shuffle=False  # time-ordered split
)

model = XGBClassifier(
    n_estimators=300,
    max_depth=5,
    learning_rate=0.05,
    subsample=0.8,
    colsample_bytree=0.8,
    use_label_encoder=False,
    eval_metric="logloss",
)
model.fit(X_train, y_train)

probs = model.predict_proba(X_test)[:, 1]
print(f"Brier score: {brier_score_loss(y_test, probs):.4f}")  # target ≤ 0.21
print(f"AUC-ROC:     {roc_auc_score(y_test, probs):.4f}")    # target ≥ 0.67
```

```python
# Plot team Elo history
import polars as pl
import matplotlib.pyplot as plt

elo = pl.read_parquet("data/team_elo/team_elo_2024.parquet")
team = elo.filter((pl.col("team_name") == "NAVI") & (pl.col("map_name") == "all"))

plt.plot(team["date"].to_list(), team["elo_before"].to_list())
plt.title("NAVI Global Elo — 2024")
plt.xlabel("Date")
plt.ylabel("Elo")
plt.show()
```

---

## Methodology Notes

### Elo model
The Elo ratings in GoldLite are computed using a custom variant of the standard Elo formula:

- **K-factor**: Variable by tier (T1 = 32, T2 = 24, T3 = 20) and match format (bo5 weighted higher)
- **Map-specific Elo**: Separate rating maintained per CS2 map (de_mirage, de_dust2, etc.)
- **Initial Elo**: All teams start at 1500. Teams with fewer than 10 matches have higher uncertainty.
- **Update frequency**: Recomputed from all historical results on demand.

The full GoldLite Elo model is a **baseline** — it does not incorporate demo-derived features, player-level signals, or market information. It is calibrated to be reproducible and interpretable, not to maximize prediction accuracy.

CounterQuant's production model (internal, not released) additionally incorporates 509 features including economy statistics, KAST, utility efficiency, and market-implied probabilities.

### No data leakage guarantee
All features in `match_features_YYYY.parquet` are constructed from data available strictly **before** the match date. The `team1_won` column is the label — it is derived from the match result after the fact. There are no post-match statistics in the feature columns.

---

## Prediction Targets & Baselines

| Model | Target | GoldLite Baseline | Production Target |
|-------|--------|------------------|-------------------|
| Match outcome | team1_won (binary) | Brier ≤ 0.22 | Brier ≤ 0.21 |
| — | — | AUC-ROC ≥ 0.65 | AUC-ROC ≥ 0.67 |
| Map outcome | map winner (binary) | Brier ≤ 0.24 | TBD |

A random classifier scores Brier = 0.25. An Elo-only model should reach approximately 0.228. The GoldLite feature set adds form, H2H, and event context on top of Elo and should reach around 0.22.

---

## Related Datasets

| Dataset | Contents | Relationship |
|---------|----------|--------------|
| [CounterQuant CS2 Demos](https://huggingface.co/datasets/KEDevO/CounterQuant-CS2-Demos) | Raw `.dem` files | Source layer |
| [CounterQuant CS2 Bronze](https://huggingface.co/datasets/KEDevO/CounterQuant-CS2-Bronze) | Tick-level events: kills, damages, flashes, utility | Compute demo-derived features to extend GoldLite |
| [CounterQuant CS2 Silver](https://huggingface.co/datasets/KEDevO/CounterQuant-CS2-Silver) | Match results, map scores, player stats | GoldLite is derived from Silver — join on `match_id` |
| [CounterQuant Platform](https://counterquant.com) | Live analytics, predictions, API | Production system built on the full 509-feature pipeline |
| [CounterQuant API](https://counterquant.com/api/docs/) | REST API access | Access to live predictions and extended features |

---

## Data Release Policy

| Layer | Public? | Timeline |
|-------|---------|----------|
| **GoldLite 2024** (Elo + form + H2H + event) | ✅ CC BY 4.0 | Q3 2026 |
| **GoldLite 2012–2023** (backfill) | ✅ CC BY 4.0 | Q4 2026 |
| **GoldLite 2025** | ✅ CC BY 4.0 | 2027 (2-year delay) |
| Demo-derived features (eco, KAST, utility) | ❌ Private | CounterQuant API only |
| Full 509-feature vectors | ❌ Private | CounterQuant API only |
| Polymarket / market odds | ❌ Private | Never released |
| Model output probabilities | ❌ Private | CounterQuant API only |

---

## Citation

If you use this dataset in research, analytics products, or publications, please cite:

### BibTeX

```bibtex
@dataset{kulbe2026counterquantgoldlite,
  author       = {Eimantas Kulbe},
  title        = {CounterQuant CS2 GoldLite: ML-Ready Match Features for Professional CS2},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/KEDevO/CounterQuant-CS2-GoldLite},
  note         = {Pre-match Elo ratings, team form, head-to-head records, and event
                  context features for professional CS2 match outcome modelling.
                  Covers 2012--2024 (first release). 2-year delay for recent data.}
}
```

### APA

Kulbe, E. (2026). *CounterQuant CS2 GoldLite: ML-Ready Match Features for Professional CS2* [Dataset]. Hugging Face. https://huggingface.co/datasets/KEDevO/CounterQuant-CS2-GoldLite

### Acknowledgement (for papers/articles)

> Pre-match features and Elo ratings sourced from *CounterQuant CS2 GoldLite* (Kulbe, 2026), available at https://huggingface.co/datasets/KEDevO/CounterQuant-CS2-GoldLite under CC BY 4.0.

---

## License

**Creative Commons Attribution 4.0 International (CC BY 4.0)**

You are free to:
- **Share** — copy and redistribute the material in any medium or format
- **Adapt** — remix, transform, and build upon the material for any purpose, including commercial

Under the condition that you **give appropriate credit** to **Eimantas Kulbe** and link back to this dataset.

Full license text: https://creativecommons.org/licenses/by/4.0/

---

*Dataset maintained by [Eimantas Kulbe](https://counterquant.com). For questions, issues, or collaboration: open a discussion on this dataset page or reach out via [CounterQuant](https://counterquant.com).*