cutrope/README.md

SimVerse / cutrope

Cut the Rope video-to-command: watch a short gameplay video of one Cut the Rope level (with a coordinate grid overlay) and output a deterministic command script that, replayed from the same initial state, wins the level.

• Records: 272 levels
• Modality: short MP4 gameplay video (~3s each, 1920×1080)
• Output: {"commands": "...", "reason": "...", "confidence": 0..1} (open-ended; many winning scripts may be valid)

Loading

from datasets import load_dataset

ds = load_dataset("SimVer-ano/simverse2026", "cutrope")
example = ds["test"][0]

system_text = example["prompt"]["system"]
user_text   = example["prompt"]["user"]

video_path = example["video_relative_to_config"]["path"]   # e.g. "videos/rope-000.mp4"
mime_type  = example["video_relative_to_config"]["mime_type"]

gold_commands = example["answer"]["commands"]              # one known-3-star reference script

Schema

Field	Type	Description
level_id	string	Level id, e.g. "rope-000"
__sample_id__	string	Same as level_id
prompt.system / prompt.user	string	Exact prompt text
video.path	string	Original (project-relative) video path
video.mime_type	string	Always "video/mp4" for shipped variants
video.duration_seconds	int	Video length
video.fps	int	Frame rate
prompt_level.canvas_width / canvas_height	int	Game canvas dimensions
prompt_level.object_counts	dict	Per-object-type counts: target, candy, star, bubble, pump, gravity_button, grab_or_rope_anchor, left_candy, right_candy
prompt_level.two_parts	bool	Whether the level starts with split-candy halves
level_json_without_solution	dict	The full level definition stripped of the gold solution; useful if you want to recreate the simulation state without seeing the answer
video_relative_to_config	dict	path / mime_type rewritten to be relative to this config's root
answer.commands	string	Reference winning command script (one command per line)
answer.reason	string	Short explanation of the reference solution
answer.confidence	float	Always 1.0 for the shipped reference
legacy_answer	string	Same script under the pre-v1 field name reference_solution (kept for back-compat)

Command DSL

Commands accept these actions, optionally guarded by when <CONDITION>:

cut_rope N | cut_rope N,M,K       # cut one or more ropes
pop_bubble N | pop_bubble_left | pop_bubble_right
activate_pump N [times C] [every S] [until <CONDITION>]
move_grab N X | move_grab N X Y
kick_rope N
toggle_gravity

Conditions include candy_x/candy_y thresholds, candy_near X,Y,R [for S], candy_still for S, rope_cut N, no_rope, candy_in_bubble, plus split-candy variants left_candy_* / right_candy_*. Boolean combinators and / or and parentheses are supported. wait_frames is not allowed in this benchmark version — use condition-based waits.

See the system + user prompt embedded in each record for the full vocabulary.

Companion files

• data/<id>.json — per-record JSON (mirror of the JSONL records as individual files).
• source/<id>.json — the original frontend-friendly level files (with textCommandSolution, gameplay metadata for the in-browser engine).
• videos/<id>.mp4 — the gameplay clips.

License

MIT — see LICENSE at the repo root.