Datasets:
IL_Replay_1115
This anonymized Clash Royale replay dataset was frozen on November 15, 2026. It contains 156,160 1v1 replays and 10,976,839 card plays and ability activations from November 4 through November 15 (UTC). The dataset supports offline gameplay analysis and imitation learning research.
Coverage
| Mode or group | Replays |
|---|---|
| Ranked · Grand Champion | 365 |
| Ranked · Royal Champion | 35,506 |
| Ranked · Ultimate Champion | 119,539 |
| Grand Challenge · winner's 10th win | 287 |
| Grand Challenge · winner's 11th win | 243 |
| Grand Challenge · winner's 12th win | 220 |
| Total | 156,160 |
The dataset includes 155,410 Ranked battles and 750 Grand Challenge battles. Both tables use the train split. When creating evaluation splits, group by replay_tag to keep both player perspectives and all actions from the same battle together.
Files and schema
The dataset structure follows VanguardX101/IL_Replay.
| Path | Contents |
|---|---|
replays/part-*.parquet |
One row per battle. Includes the mode, result, crowns, event count, ranked league or challenge win number, and full payload_json. |
actions/part-*.parquet |
One row per event. Join on replay_tag and use source_index to recover the order of the source markers within a replay. |
manifest.json, SHA256SUMS |
Coverage, anonymization, validation and SHA-256 checksums for all 64 Parquet shards. |
LICENSE |
Apache License 2.0 text for the original contributions described below. |
The replay record preserves both players' decks and displayed card levels, tower troops and their levels, final tower hit points, crown counts, the result, replay-relative duration where available, and all recorded actions.
Actions use a 20 Hz clock.
Loading the data
Run the following from the local dataset directory.
import json
from datasets import load_dataset
replays = load_dataset("parquet", data_files="replays/*.parquet", split="train", streaming=True)
actions = load_dataset("parquet", data_files="actions/*.parquet", split="train", streaming=True)
row = next(iter(replays))
payload = json.loads(row["payload_json"])
assert payload["source"]["replay_tag"] == row["replay_tag"]
assert len(payload["events"]) == row["event_count"]
Adapting to FirstLight engines
FirstLight users need an adapter for their engine version. Check which fields the target version accepts and how it validates them. The original IL_Replay importer expects royaleapi-battle-actions.v1, whose fields and validation counts differ from those in this dataset.
- Engine and resource versions. Compatibility was checked on a sample using client version code
160635007and resource baseline16.635.8-062a11e0. Check that the client and resources match the target runtime version and host architecture. Resolve card forms, tower troops, mode IDs and level caps using the target engine's catalogs. - Code and execution semantics. Use
data_ito mapteamandopponentto engine sides. Check coordinate conversion, source action order, the meaning of 20 Hz ticks, the timing of input submission and action execution, and how the engine maps an ability source to a runtime object. Handle the card's deck form and its form at play separately. - Initialization and validation. Calibrate a candidate opening and random seed that the engine can replay. Verify executed actions, the result, crown counts and final tower hit points before exporting a training cache. Report missing fields and rejected replays in the target engine's validation format.
Anonymization and limitations
Replay IDs are random UUID4 values scoped to this release. Player labels player-1 and player-2 reset each match. Original identifiers, links, exact battle and collection times, raw HTML, source paths and per-replay source hashes were removed. Someone with the original action sequence may still be able to match the battle. The dataset contains discrete actions and final results. Original random seeds, opening hands, card cycles and continuous battle states are absent.
License and provenance
The original documentation and licensable compilation work are available under Apache License 2.0.
- Downloads last month
- -