erickfm/melee-ranked-replays

Melee Ranked Replays

Anonymized Slippi ranked replays (platinum+) from Super Smash Bros. Melee, sharded by character and rank pair. Built for behavior-cloning and other replay-driven ML work on Melee — notably MIMIC.

Contents

Raw .slp files grouped into tarballs by (character, rank_pair, source_archive):

shards/
  {CHAR}_{rank_pair}_a{N}.tar.gz
  metadata_a{N}.json

• Characters: 25 distinct (ZELDA and SHEIK collapsed to ZELDA_SHEIK; POPO and NANA collapsed to ICE_CLIMBERS)
• Rank pairs: diamond-diamond, diamond-platinum, master-diamond, master-master, master-platinum, platinum-platinum (6 combos, higher rank first in mixed pairs)
• Source archives: a1..a6, corresponding to the 6 original anonymized ranked dumps. Archive suffix exists so incremental uploads don't collide; if you want "everything Fox at master-master" you pull every FOX_master-master_a*.tar.gz.

Each shard holds the raw .slp files — no preprocessing, normalization, or tensorization applied. Use peppi-py, py-slippi, or libmelee to parse.

Duplication

Each replay is placed into both players' character buckets (unless it's a ditto). A MARTH vs FALCO diamond-platinum replay appears in both MARTH_diamond-platinum_aN.tar.gz and FALCO_diamond-platinum_aN.tar.gz. A FOX ditto only appears once in FOX_diamond-diamond_aN.tar.gz. This means downloading "all Marth games at master-master" needs only a single tarball (not a join across 25 per-player files), at the cost of ~90% duplication on the full dataset.

Metadata

shards/metadata_a{N}.json is a flat JSON list. One entry per replay (not per bucket), schema:

{
  "filename": "diamond-diamond-6cf8c1ee745993cefe0c88db.slp",
  "p1": "NESS",
  "p2": "JIGGLYPUFF",
  "rank": "diamond-diamond",
  "archive": "3"
}

p1 and p2 use the same collapsed character names as the bucket filenames. archive is a string.

Build pipeline

Source: six anonymized ranked archives covering ~850k total replays at platinum+ rank. Each archive is processed independently by tools/shard_and_upload_ranked.py in the MIMIC repo.

Per-file work (parallel, one worker per CPU)

For each .slp file in an archive:

1. Read header only via peppi_py.read_slippi(path, skip_frames=True) — skipping frames makes it fast (ms per file) since we only need the Start event, not the ~10k frames of gameplay.
2. Pull the 2 players out of game.start.players, reject if not exactly 2.
3. Map each player's character int to a name via a lookup built from melee.Character enum, with two collapses:
   • ZELDA (19) and SHEIK (7) → ZELDA_SHEIK (same fighter mid-match)
   • POPO (10) and NANA (11) → ICE_CLIMBERS (two climbers are one unit)
4. Reject junk characters: WIREFRAME_MALE/FEMALE, GIGA_BOWSER, SANDBAG, UNKNOWN — not legal tournament characters; replays featuring them are debug/test files.
5. Parse rank from filename via regex — the {rank1}-{rank2} prefix.

Per-file output: (filename, p1_name, p2_name, rank_pair, error_or_None).

Bucketing

Each successful replay enters up to two buckets keyed by (character, rank_pair):

• One for player 1's character
• One for player 2's character (skipped if same char — no double-counting dittos)

Metadata is a flat list of {filename, p1, p2, rank, archive} entries, one row per replay.

Tar + upload

Buckets are compressed one at a time (tarfile w:gz, compresslevel=6), uploaded via huggingface_hub.HfApi.upload_file, then the local tar is deleted. Already-uploaded paths are skipped, so the tool is resume-safe.

Intended use

Training behavior-cloning models on human gameplay (MIMIC, HAL, and similar projects), replay analysis, frame-data research. If you're building something on this, drop me a line.

License

MIT. Replays were originally published by the Slippi/ranked community in anonymized form; this is a re-sharded redistribution for ML convenience.