elliottash/megaman-maker-levels

Mega Man Maker level corpus

A unified corpus of 227,604 user-authored Mega Man Maker levels for training MarioGPT-style text-to-level models, procedural-content generation research, and platformer-level analysis.

Each row carries:

• A MarioGPT-style ASCII tile grid (level_ascii) — list[str], one string per row of the level — ready to feed directly into a GPT2 / BPE tokenizer with no further preprocessing.
• A natural-language text_prompt automatically composed from the level's structural features ("underwater, many enemies, few pickups, has boss, medium, low elevation, difficulty expert") — modeled after MarioGPT's prompt vocabulary.
• The full parsed intermediate representation (IR) as gzipped JSON (ir_json_gz).
• The original raw .mmlv bytes (mmlv_bytes) when retained.
• The full webmeka.io engagement metadata — likes, dislikes, downloads, crowd-sourced difficulty, boss enum, author, creation timestamp.
• A webmeka_url link back to the level's webmeka.io page.

Quickstart

from datasets import load_dataset

ds = load_dataset("<namespace>/<repo>", "default", split="train", streaming=True)
ex = next(iter(ds))
print(ex["text_prompt"])
#> underwater, some enemies, few pickups, has boss, medium, medium elevation
print("\n".join(ex["level_ascii"][:14]))

Train MarioGPT-style with "\n".join(level_ascii) as the GPT2 input and text_prompt as the conditioning prompt.

Configs / subsets

Config	Rows	Description
default	227,604	Every parseable level.
mmm_liked	55,124	likes>=10 AND quality_score>=0.5 — training-grade.
mmm_boss	201,285	Levels containing an authored boss entity (boss_entity_count > 0; boss doors alone do not qualify).
mariogpt_compat	30,927	ASCII grid ≤17 rows tall and ≤256 cols wide.

ds = load_dataset("<namespace>/<repo>", "mmm_liked", split="train")

ASCII legend (level_legend_version = "v2")

Glyph	Meaning
-	empty / passable
X	solid (ground / wall / breakable)
=	semisolid / one-way / jump-through
H	ladder
W	water / liquid zone
*	hazard (spike, lava surface)
?	item block (solid + collectible)
o	pickup / collectible
E	enemy
B	boss (an actual boss entity)
D	boss door (new in v2)
G	goal / level exit
P	player spawn
g	gizmo (switch, spring, pipe, ...)
~	slippery solid (reserved; new in v2)
.	unknown / reserved

Legend v1 → v2: boss doors were folded into B in v1, which made a B appear in 92% of grids; in v2 B marks a real boss fight and doors render as D. `` is reserved for slippery (ice) tiles — currently unobserved in source data. Grids generated for v1 remain retrievable from this repo's git history.

The mmm_decode.py helper that ships in this repo includes a Pillow-based ASCII-to-PNG renderer using the same palette.

Column schema

Column	dtype	Description
mmm_id	int64	Numeric Mega Man Maker level id (matches webmeka.io URL).
webmeka_url	string	https://webmeka.io/level/{mmm_id}.
ir_id	string	mmm_<mmm_id>; matches Meta.id in the parsed IR.
level_ascii	list	MarioGPT-style tile grid, one string per row.
level_legend_version	string	v2. Increments when the glyph table changes.
split	string	Official split (train/validation/test, 98/1/1). Author-disjoint: all levels by one author share a split; exact byte-duplicates are co-assigned. Use it to avoid author/duplicate leakage.
ascii_width	int64	Width of the ASCII grid in tiles.
ascii_height	int64	Height of the ASCII grid in tiles.
text_prompt	string	Comma-separated MarioGPT-style prompt.
theme	string	First useful source tag (platformer fallback).
elevation	string	low elevation / medium elevation / high elevation.
width_bucket	string	single-screen / short / medium / long / very long.
height_bucket	string	low ceiling / medium ceiling / tall.
enemy_count	int64	Count of E glyphs.
pickup_count	int64	Count of o + ? glyphs.
hazard_count	int64	Count of * glyphs.
gizmo_count	int64	Count of g glyphs.
ladder_count	int64	Count of H glyphs.
water_count	int64	Count of W glyphs.
boss_count	int64	Count of B glyphs (real boss entities under legend v2; no webmeka inflation).
boss_door_count	int64	Count of D glyphs (boss doors).
boss_entity_count	int64	Boss entities in the IR (d=8 cells that are neither doors nor goal capsules). Drives the mmm_boss config.
boss_door_entity_count	int64	Boss-door entities in the IR.
boss_names	list	Human-readable names of authored bosses (from the game's data.win tables), e.g. bubble_man.
spawn_count	int64	Count of P glyphs.
goal_count	int64	Count of G glyphs.
has_ladders	bool	True when ladder_count > 0.
has_water	bool	True when water_count > 0.
has_boss	bool	True when the grid contains a boss entity (boss_count > 0).
glyph_counts	struct<string,int64>	Glyph histogram across the entire legend (0 for absent glyphs, so every row has the same key set).
webmeka_name	string	Level title as displayed on webmeka.io.
webmeka_author_id	int64	Webmeka author id.
webmeka_author_name	string	Webmeka author display name at scrape time.
webmeka_author_icon	int64	Author icon enum.
webmeka_created	string	ISO-8601 creation timestamp.
webmeka_boss	int64	Boss-portrait enum from webmeka. Nearly every level sets it (99.996% nonzero) — do not use it as a boss indicator; use boss_entity_count.
webmeka_likes	int64	Number of likes.
webmeka_dislikes	int64	Number of dislikes.
webmeka_downloads	int64	Number of downloads.
webmeka_difficulty_score	float64	Crowd-sourced difficulty in [0, 1].
webmeka_difficulty_label	string	Bucketed difficulty (Easy, Normal, Hard, Expert, Super Expert).
webmeka_tags	list	Webmeka tags (traditional, puzzle, boss, automatic, …).
name	string	Name as recorded inside the .mmlv.
author	string	Author as recorded inside the .mmlv.
difficulty_label	string	Same 5-bucket scheme as webmeka_difficulty_label (shared bucket_difficulty function since parser v5; null when the level is unrated).
ir_quality_score	float64	likes / (likes + dislikes).
width_tiles	int64	Level width in 16×16 tiles.
height_tiles	int64	Level height in 16×16 tiles.
tile_width_px	int64	Source tile width in px (always 16).
tile_height_px	int64	Source tile height in px (always 16).
tag_list	list	Tags carried inside the IR (megaman_maker, platformer).
ir_format_version	string	Parser format version (mmm_parser_v5).
region_count	int64	Number of raw MMM regions (one per contiguous tileset block).
region_tileset_ids	list	Raw MMM tileset_id per region.
screen_count	int64	Number of webmeka screens stamped in extra.screens (nullable for legacy IR).
h_transition_count	int64	Screens with a nonzero horizontal-transition setting.
v_transition_count	int64	Screens with a nonzero vertical-transition setting.
is_underwater	bool	Any region/screen uses an underwater tileset ({5, 29, 56}).
playable_screen_count	int64	Screens in the authored playable mask (nullable for legacy IR).
playable_area_ratio	float64	Playable-mask area over the IR bounding box (≤1.0; nullable).
spike_tile_count	int64	Tiles stamped with a directed spike (v5 IR).
spike_dir_counts_json	string	JSON map raw-MMM-spike-orientation-enum → tile count.
setting_sliding	bool	Slide ability enabled (nullable when absent from the .mmlv).
setting_charge_shot	bool	Charge shot enabled (nullable).
setting_double_jump	bool	Double jump enabled (nullable).
setting_double_damage	bool	Double damage enabled (nullable).
setting_proto_strike	bool	Proto strike enabled (nullable).
weapon_ids	list	Raw weapon-slot ids (1k0..1k11).
weapon_count	int64	Nonzero weapon slots.
music_level_id	int64	Level-music id (nullable).
music_level_game	int64	Level-music source-game enum (nullable).
music_boss_id	int64	Boss-music id (nullable).
boss_portrait_id	int64	The boss-portrait enum recorded inside the .mmlv (cosmetic).
boss_weakness_items	list	Enabled boss-weakness item ids.
default_bg_color	int64	Default background-color enum (nullable).
total_entities	int64	Entity count across all layers / regions.
first_tileset	string	Name of the first IR tileset (rough biome hint).
tile_layer_count	int64	Number of top-level tile layers.
tileset_count	int64	Number of distinct tilesets referenced.
ir_sha256	string	SHA-256 of the canonical-JSON IR bytes (for dedupe).
ir_json_gz	binary	gzipped UTF-8 JSON of the full IR Level.
mmlv_bytes	binary	Raw .mmlv file content (when retained).
mmlv_sha256	string	SHA-256 of the raw .mmlv bytes (for dedupe).
thumbnail_png	binary	Pre-rendered PNG of the ASCII grid (when --thumbnails was used).

Provenance & license

• Source content — user-authored Mega Man Maker levels hosted on webmeka.io. The .mmlv files contain ONLY symbolic ids (e.g. d=8 e=22 means "Robot Master id 22" — Bubble Man inside the Mega Man Maker engine) plus tile coordinates and screen settings. They contain no Capcom assets — no sprites, no music, no images — only integer references to assets the Mega Man Maker engine renders at play time.
• Metadata — webmeka_* columns mirror the public webmeka.io search API as observed on 2026-07-03.
• Parser — level_ascii, text_prompt, and ir_json_gz are derived by the open-source level-corpus pipeline (format version mmm_parser_v5).

The level layouts are user-authored, and this dataset is released under CC-BY-4.0 for research and PCGML use. "Mega Man" is a trademark of Capcom Co., Ltd.; this dataset is not affiliated with or endorsed by Capcom. If you are a level author and would like your levels removed, please open an issue on the dataset repo.

Quality filtering

The default config is unfiltered so consumers can apply their own quality bar. The included engagement columns make this easy:

ds = ds.filter(
    lambda x: x["webmeka_likes"] >= 10
    and x["webmeka_dislikes"] < x["webmeka_likes"]
    and x["webmeka_downloads"] >= 20
)

Or load the precomputed mmm_liked config directly.

Distributions (this build)

• Total rows: 227,604
• Rows carrying raw .mmlv: 227,604
• Rows carrying parsed IR: 227,604
• Rows carrying ASCII grid: 227,604
• Rows carrying thumbnail PNG: 227,604
• IR format version breakdown: mmm_parser_v5 (227,604)
• Difficulty buckets: Hard (50,251), Expert (40,306), Normal (38,122), Super Expert (25,922), Easy (19,384)
• Top themes: data_archives (89,362), platformer (53,763), factory (25,920), castle (21,048), underwater (9,260), forest (9,220), snow (6,408), fortress (6,284), desert (4,282), ghost_house (2,057)

Known exclusions

Levels whose active-tile bounding box is smaller than 4×4 tiles (degenerate strips, ~420 of them) are intentionally excluded: they are not playable layouts. Their raw .mmlv files exist locally but produce no IR and therefore no row.

Version history

• v2 (this build) — parser mmm_parser_v5, ASCII legend v2, ~50K additional levels (completed catalog crawl + refreshed engagement metadata). Fixes over v1:
  • Boss labeling: v1's boss_count/has_boss/mmm_boss config keyed on the webmeka boss-portrait enum and on a B glyph that included boss doors, so ~100% of rows counted as "has boss". v2 counts actual boss entities (boss_entity_count); boss doors render as D.
  • Semisolid platforms: v1 IR (parser v2) contained zero one_way tiles — jump-through platforms were silently dropped. Restored via the game's own object→affordance tables.
  • Regions: v1 rows all had region_count = 0; v2 stamps contiguous same-tileset screen regions (extra.regions) and region_tileset_ids.
  • Spike orientation: preserved per-tile (spike_dir_counts_json, IR TileInfo.properties.mmm_spike_dir); v1 collapsed all spikes to one undirected tile.
  • ON/OFF blocks: now stamp a solid tile footprint tagged on_off_red/on_off_blue; v1 left them collision-less.
  • Difficulty: difficulty_label and webmeka_difficulty_label now share one 5-bucket scheme (v1 used two incompatible schemes).
  • New columns: split, entity names, ability/weapon/music settings, playable-mask stats, is_underwater, transitions.
  • v1 files remain retrievable from this repo's git history.
• v1 (2026-05-23) — initial release: 179,166 levels, parser mmm_parser_v2, legend v1.

Citation

If you use this corpus, please cite the level-corpus repository (which ships the .mmlv parser used to produce the IR + ASCII + text prompts) along with the webmeka.io community as the data source.

Entity behavioral taxonomy (entity_categories config, added v2.1)

Every distinct MMM entity (367: enemies, bosses, pickups) is labeled with a behavioral family — WALKER, FLYER, SHOOTER, JUMPER, CLIMBER, AQUATIC, STATIONARY, CHARGER, BURROWER, BOSS, ENEMY_OTHER — grouped under a coarse entity_class (ENEMY / PICKUP / GIZMO / ...). Load it with load_dataset("elliottash/megaman-maker-levels", "entity_categories") and join to level entities on the object slug.

Every label carries its provenance, assigned by a strict hierarchy so the labels are auditable rather than asserted:

provenance	source	count
ground_truth	game binary — d-channel class, data.win prtMidBoss parent for mid-bosses	108
wiki	read from the Mega Man Maker wiki / Mega Man fan wikis; source_url + evidence recorded per row	254
llm / kb_llm	model knowledge, only where no wiki page was found (flagged; 5 rows)	5

Columns: fine, slug, display_name, object_name, kind, entity_class, family, provenance, confidence, source_url, evidence, parent_class, wiki_family. wiki_family preserves the wiki guess where a data.win parent class overrode it (6 enemies the wiki mislabeled by their attack were corrected to BOSS via the prtMidBoss parent). parent_class is the object's GameMaker parent from data.win. Filter on provenance/confidence for a high-precision subset; the 3 low-confidence (<0.6) rows are edge cases (a spike-enemy and two spawners). Note: MMM encodes movement in event code, not the class hierarchy, so parent classes confirm boss/spawner status but do not by themselves give walker-vs-flyer.