---
configs:
- config_name: default
  data_files:
  - split: top
    path:
    - top/*.parquet
    - days_0721_0731/top/*.parquet
  - split: medium_high
    path:
    - medium_high/*.parquet
    - days_0721_0731/medium_high/*.parquet
size_categories:
- 100K<n<1M
tags:
- game-logs
- pokemon-tcg
- agents
---

# Pokémon TCG Tournament Game Logs

Full game logs from an agent Pokémon TCG tournament ("Limited Card Battle"),
122,542 recorded games collected across 26 daily archives — an initial batch
of 15 undated archives (`top/`, `medium_high/`) plus dated days
2026-07-21 → 2026-07-31 (`days_0721_0731/`). Both batches use the same
per-day split rule and load together via the split config above.

## Splits

Games in each daily archive were ranked by `avg_score` (per-agent average
rating, ELO-like, ~1000–1230, higher = better):

| Split | Games | Contents |
|---|---|---|
| `top` | 18,200 | the 700 highest-rated games from each archive (high-quality / eval set) |
| `medium_high` | 104,342 | every other game (pretraining set) |

Episode IDs are globally unique across archives; the two splits are disjoint.

## Columns

| Column | Type | Description |
|---|---|---|
| `episode_id` | int64 | unique game/episode id (original JSON filename) |
| `agents` | list\<string\> | the two agent names (`info.TeamNames`) |
| `rewards` | list\<double\> | final rewards per agent, e.g. `[1, -1]` (win/loss), `[0, 0]` (draw); may be null |
| `json` | string | the complete original game record, verbatim JSON |

Each `json` value is one full game dump (~2–9 MB raw) with top-level keys:
`configuration, description, id, info, module_version, name, rewards,
schema_version, specification, statuses, steps, title, version` — including
the complete step-by-step game trajectory under `steps`.

## Usage

```python
from datasets import load_dataset
import json

ds = load_dataset("shantezhou/pokemon_data", split="top", streaming=True)
for row in ds:
    game = json.loads(row["json"])
    print(row["episode_id"], row["agents"], row["rewards"])
    break
```

Parquet shards are zstd-compressed (~80x vs raw JSON); the full dataset is
~300 GB of raw JSON stored in a few GB of parquet.