smearle/puzzlescript-gists

PuzzleScript Human-Authored Games (Deduplicated)

A corpus of 23,667 human-authored PuzzleScript games — the complete source text of each game — after structural deduplication and removal of non-vanilla (PuzzleScript Plus) dialects. The corpus is intended as training material for world models of grid-puzzle dynamics, where each game defines a set of rewrite rules (the mechanics) and one or more levels (initial states), but it is equally usable for code generation, program synthesis, or game-design research.

Dataset structure

One JSON object per game in data/puzzlescript_games.jsonl:

field	description
id	game identifier (original file stem)
content	full PuzzleScript source text
provenance	curated_collection or gist_scrape (see Data collection)
parse_status	ok if it parses under a vanilla PuzzleScript grammar; otherwise the failure class
n_objects	number of canonical objects (parsed games only)
n_levels	number of levels (parsed games only)
mechanics_hash	name/color/sprite-invariant fingerprint of the rules (parsed games only)
levels_hash	name-invariant fingerprint of the level layouts (parsed games only)
n_duplicates	number of near-identical copies collapsed into this entry
duplicate_ids	the collapsed copies' identifiers

Two manifests accompany the data:

• dedup_master.json — the full dedupe record: every kept game, its grouping key, and every duplicate collapsed into it, so any cluster can be reconstructed.
• quarantine_ps_plus_manifest.txt — the identifiers of the 2,357 PuzzleScript Plus files removed before deduplication.

Data collection

The corpus is built from two kinds of source, all originally public:

• A base of ~26,000 PuzzleScript games that had previously been scraped from GitHub Gist by a third party. These carry provenance: gist_scrape.
• Several curated public PuzzleScript collections (the official game gallery and other openly published sets), merged in and marked provenance: curated_collection. Where a game appeared in both a curated set and the gist base, the curated copy is preferred.

Files were collected as raw .txt source. No emulation, rendering, or playthrough data is included — only the game programs themselves.

Filtering: PuzzleScript Plus removal

PuzzleScript Plus (PS+) is a superset dialect with extra TAGS / MAPPINGS sections and additional prelude directives (level_select, sprite_size, mouse_*, tween_*, case_sensitive, …) that a vanilla PuzzleScript engine cannot run. We detect PS+ with positive whitelists — the PS+-only section headers and prelude keywords — rather than "anything not vanilla," which would over-flag typos and prose. 2,357 files were flagged; 98.8% of them independently failed a vanilla PuzzleScript parser, confirming the detector and the parser agree. These files were removed before deduplication (their identifiers are listed in quarantine_ps_plus_manifest.txt). The whitelist is high-precision but not exhaustive: a PS+ game that uses only extended rule semantics — with no PS+ section or prelude directive — can look vanilla and may remain in the corpus.

Deduplication procedure

We treat a game as a duplicate only when it offers no novel training material — i.e. it matches another game in both mechanics and levels. Any difference in rules or in level layouts is preserved. Deduplication runs as a layered pipeline, cheap to expensive, each stage only refining the survivors of the last:

1. Exact — hash of the raw bytes.
2. Normalized — whitespace-collapsed, case-folded hash.
3. Lite — the game is run through the engine's own preprocessing (comment stripping with full nesting support, sound/message removal, whitespace and section-header normalization), then the mechanics sections and the levels are hashed separately. Name- and art-sensitive; no parser.
4. Canonical tokens — the game is parsed and tokenized into a name-, color-, and sprite-invariant representation, split into a mechanics fingerprint and a levels fingerprint. Recolored sprites, redrawn art, and renamed objects collapse here; functional prelude directives that the tokenizer would otherwise drop are folded back into the mechanics fingerprint so they still count as a difference.

A game's dedupe identity is the pair (mechanics_hash, levels_hash) from the deepest stage it reaches. Unparseable games fall back to the lite or exact fingerprints in separate namespaces, so a game that could not be parsed is never silently merged into a parsed one.

Statistics

	count
Collected (raw)	35,075
PuzzleScript Plus removed	2,357
Vanilla games deduped	32,718
Published (deduped)	23,667
— parsed to canonical tokens	21,994
Distinct mechanics fingerprints	15,960
Distinct level-set fingerprints	18,841

Among parsed games, 2,622 mechanics fingerprints are shared by two or more distinct games — the "same dynamics, novel levels" cohort that the dedupe policy deliberately keeps.

Intended use & limitations

• Intended use: training and evaluating world models / dynamics models of PuzzleScript, plus code-generation and game-design research.
• Parse failures: ~1,673 published games do not parse under a vanilla grammar (malformed source, encoding quirks, or undetected PS+ rule syntax). They are kept because they may still be valid for some engines and because the lite fingerprints still deduplicate them, but they lack the mechanics_hash / levels_hash fields.
• Deduplication is structural, not semantic. Two games with logically equivalent but textually different rules are treated as distinct.

Authorship & licensing

These are community-authored works gathered from public sources. Individual games carry their own authorship and licensing terms, which are not uniform and are not asserted here (license: other). This collection is released to support research and reproducibility; please respect the original authors and their licenses when reusing individual games. If you are an author and want a game removed, please open a discussion on this repository.