JonesLin/next-jev-tetris-decisions-10k

Tetris placement decisions with known answers (10k)

10,000 Tetris decisions for collecting reasoning from an LLM teacher: the teacher writes its reasoning and a final \box{X}, and verify.py keeps only answers that match the known best option.

Each item is exactly what the community Jev Tetris game (commit a0b411c) sends to a local model in its "Laya" slot: a one-paragraph board description, the instruction, and the game heuristic's top 6 placements. Option order is shuffled and relabeled A to F.

Layout

Path	Give to the teacher agents?	Content
tasks/PROMPT.md	yes	Instructions for each agent
tasks/batches/batch-0000.jsonl ... batch-1249.jsonl	yes	8 items per batch, no answers
tasks/items.jsonl	yes	All 10,000 items, no answers
grading/answers.jsonl	no	{id, answer}
grading/meta.jsonl	no	Seed, piece, heuristic score of every option, text_tie, game option ids
verify.py	no	The verifier
tools/gen_decisions.mjs	no	Generator (imports the game engine; see below)

Put only tasks/ in the agents' workspace. grading/ holds the answers.

Item fields: id, state, instructions, options (A-F to text), prompt (all of it as one text). Example prompt:

Tetris. The stack is low and slightly uneven with several holes. The falling piece is an S. Next piece: T.

Question: Pick the best placement: clear lines, avoid holes, keep the stack low and flat.
Options:
A. columns 8-9: no lines cleared, no holes, stack grows by one row
B. columns 1-2: no lines cleared, creates one hole, surface bumpy
C. columns 3-5: no lines cleared, creates one hole, surface slightly uneven
D. columns 2-3: no lines cleared, no holes, surface bumpy
E. columns 8-10: no lines cleared, creates one hole, stack grows by one row
F. columns 4-5: no lines cleared, creates one hole, surface bumpy

Collect and verify

Each agent reads one batch and writes outputs/<id>.md per item: short reasoning, then \box{X} as the last line (see tasks/PROMPT.md). Then:

python verify.py --outputs outputs --out kept.jsonl

An output is kept only when its last \box{X} (or \boxed{X}) ends the file, the reasoning before it is not empty, and X equals the answer. kept.jsonl rows are the item plus answer and reasoning (text before the box). The summary prints counts per failure type and the kept rate for clear and tie items.

How the data was made

• The 10,000 items come from 2,296 self-play games (seeds from 100000; our Tetris evaluation uses seeds 42-46). Each game mixes the heuristic move with a random move at rate 0.05, 0.2, 0.4 or 0.7, so boards range from calm to messy. About 1 in 5 states is sampled.
• Answer: the option with the best score from the game's own heuristic (-0.51*aggregate height + 0.76*lines - 0.36*holes - 0.18*bumpiness).
• Dropped: best minus second-best score < 0.3; the best option's outcome text equals another option's; another option visibly beats the best (more lines with no more holes, or same lines with fewer holes); duplicates.
• text_tie (in grading/meta.jsonl): another option has the same lines cleared and holes created as the answer, so only the height/surface wording separates them. 4,260 of 10,000 items (42.6%). Expect a lower pass rate on these.
• Answer letters are balanced (A-F each about 1/6). Pieces: all 7. Stack height: low 2,883, medium 2,687, high 1,651, dangerously high, close to the top 1,418, very low 1,361.

Regenerate: clone the game at a0b411c next to tools/gen_decisions.mjs as jev-tetris/, then node gen_decisions.mjs out 10000 100000 0.3 (Node 20+).

The game engine is MIT licensed; it is imported, not copied here.

Choice-format training data (choice/)

Teacher: grok-4.7 (outputs in teacher/pilot-grok-4.7/outputs; 9,185 of 10,000 verified correct). Converted with scripts/tetris_choice_data.py from github.com/JonnesLin/next_jev (commit 883e5dd) into Jev-style choice records (state, one choice question whose criteria keys are the game's placement ids, gold, teacher_rationale) for the next JEV choice head.

• Split by game: 115 of 2296 games (5%, seed 0) are held out.
• choice/train.jsonl: 8,645 items from the other games whose teacher answer was correct; the teacher reasoning is the CoT target (754 wrong-answer items dropped).
• choice/eval.jsonl: all 601 items of the held-out games (540 with teacher reasoning, the rest without), so eval is not limited to items the teacher found easy.
• choice/manifest.json: counts and SHA256 of inputs and outputs.