josefbednar/alphaclash-replays

AlphaClash replay corpus

Screen recordings of Clash Royale matches from Royale TV, scraped from a phone at 4x replay speed, each with a JSON sidecar and the deck screenshot its decks were read from. Collected for the AlphaClash agent (https://github.com/bednarjosef/AlphaClash) — a behaviour-cloned decision transformer over these replays.

Layout

manifest.jsonl                       one row per replay (see below)
replays/shard-0000/<stem>.mp4        the recording, 4x, 720x1600, ~37 MB
replays/shard-0000/<stem>.json       sidecar
replays/shard-0000/<stem>_deck.png   deck screenshot (absent on six early replays)

Shards hold 500 replays each and never change once written; new replays append to the last shard, then open the next. <stem> is <YYYYMMDD>-<HHMMSS>_<16-bit fingerprint>.

Sidecar

arena (pool: arena<N> or ranked_<league>), decks.player0/player1 (eight {slot, card, conf, level} in register-slug vocabulary), crowns [p0, p1] and winner (0/1) where read, crowns_source (card when read off the Royale TV card, tail/<rule> when read off the retained video tail, absent when null), speed_multiplier (4 — divide video timestamps by it), speed_change_at_s, battle_start_s, battle_end_s, duration_s, capture_fps, device_profile.

manifest.jsonl

stem, shard, arena, crowns, crowns_source, winner, duration_s, speed_multiplier, decks (slugs only), has_deck_png — enough to pick a subset without listing thirty thousand files.