ericyuegu/hal-mds-dataset

HAL MDS Dataset

Super Smash Bros. Melee gameplay extracted from Slippi replays. Each row is one 1v1 game: the per-frame game state and both players' controller inputs. HAL trains its Melee policies on this data.

About 1.3 million games and 13.5 billion frames (about 62,000 hours of play at 60 frames per second).

Source	Games (train / val / test)	Frames
ranked-anonymized-1 … ranked-anonymized-6	821,525 / 8,281 / 8,211	~8.9 B
pro/<player> (38 players)	473,845 / 4,878 / 4,784	~4.6 B

Layout

ranked-anonymized-1/ … ranked-anonymized-6/
pro/<player>/
player_codes.jsonl.gz

Each source folder holds:

• train/, val/, test/ — MosaicML Streaming (MDS) shards, zstd-compressed, with an index.json.
• manifest.jsonl — one line per game: replay hash, Slippi version, stage, characters, frame count, outcome, and split.
• projection.json — column names and types.
• stats.json — mean, variance, min, and max of each float feature on the train split, for normalization.
• selection.json — the filter that selected the games, with the counts it kept and rejected.
• rejections.jsonl — the games the filter rejected, and why.

Sources

ranked-anonymized-N are the six anonymized Slippi ranked-ladder replay dumps. Their names are the dump number, not a skill tier. Anonymization replaces each player's name with their tier, so p1_rank and p2_rank are 1 (Platinum), 2 (Diamond), or 3 (Master).

pro/<player> are netplay and tournament replays from 38 professional and top-level players, one folder per player. The folder's player has rank 4 (Pro) and can be on either port; the opponent has rank 0 (unknown). Some games have rank 0 on both sides, where the replay did not identify which port was the folder's player.

Filter

Every game:

• is a 1v1 between two human players with 4 starting stocks;
• is on a tournament-legal stage: Final Destination, Battlefield, Pokémon Stadium, Dream Land, Fountain of Dreams, or Yoshi's Story;
• lasts at least 1,500 frames (25 seconds);
• has at least 3 deaths and at least 100% total damage dealt and taken, and each player deals more than 50%;
• has at most one cheap death per player (a death at 10% damage or less).

Columns

Per game:

Column	Type	Meaning
replay_id	str	Replay identity; join key for manifest.jsonl and player_codes.jsonl.gz
num_frames	int	Frame count
stage	int	Stage ID
p1_character, p2_character	int	libmelee internal character ID
p1_rank, p2_rank	int	0 unknown, 1 Platinum, 2 Diamond, 3 Master, 4 Pro
p{1,2}_nana_present	int	1 if the player is Ice Climbers with Nana

Per frame, for each player (p1_, p2_) and for Nana (p1_nana_, p2_nana_):

Column	Type	Meaning
position_x, position_y	float32	Position
percent	float32	Damage
shield	float32	Shield strength
hitlag_left	float32	Frames of hitlag left
state	uint32	Action state

Per frame, each player's controller:

Column	Type	Meaning
main_stick_x, main_stick_y	int8	Main stick
c_stick_x, c_stick_y	int8	C-stick
trigger_l, trigger_r	uint8	Analog triggers
buttons	uint8	Button bitmask

Per frame, up to four items (item0_ … item3_):

Column	Type	Meaning
item_present	uint8	Bitmask of the occupied item slots
itemK_pos_x, itemK_pos_y, itemK_vel_x, itemK_vel_y	float32	Position and velocity
itemK_meta	uint32	Packed item type, state, and owner

Missing float values are NaN. HAL's hal/data/policy_world_schema.py decodes itemK_meta, and hal/data/policy_schema.py decodes the controller encodings.

player_codes.jsonl.gz

Slippi connect codes for the players in the pro/ games. The first line is a header with the code vocabulary (codes, display_names) and counts. Each following line is [replay_id, p1_player_id, p2_player_id]. Player ID 0 means the code is missing or absent from the train vocabulary. IDs from 4 upward index codes: ID 4 + i is codes[i].

Load

from huggingface_hub import snapshot_download
from streaming import StreamingDataset

root = snapshot_download(
    "ericyuegu/hal-mds-dataset",
    repo_type="dataset",
    allow_patterns="ranked-anonymized-1/val/*",
)
ds = StreamingDataset(local=f"{root}/ranked-anonymized-1/val", shuffle=False)
game = ds[0]
print(game["replay_id"], game["num_frames"], game["p1_position_x"][:10])

Credits

Thanks to Fizzi for the anonymized Slippi ranked replays (ranked-anonymized-N), and to xpilot and the professional players for the professional replays (pro/). Melee is © Nintendo / HAL Laboratory.