odats/breakout-ram4-diverse-v2

Breakout RAM4: diverse initial states, v2

A deterministic, consequence-selected benchmark for encoder–reasoner–decoder (ERD) research, with full trajectories and four-frame RAM histories. This is a new dataset, not a replacement for odats/breakout-ram4-1024. No model was trained or evaluated to select this release.

2,048 programs, 20,480 replays, 8,451,072 transitions, and 1,664 unique initial states.

Split	Programs	Replays	Initial states	Action length	Mean programs/state
train	1024	16384	1152	256–512	14.22
random_matched	256	1024	128	256–512	8.00
random_long	256	1024	128	576–768	8.00
policy_matched	256	1024	128	256–512	8.00
policy_long	256	1024	128	576–768	8.00

Training uses 1024 fitting starts and 128 validation starts, instead of 15+1 shared starts in v1. Each training program has 15 fitting replays and 1 validation replay.

Detailed report: STATISTICS.md.

What changes from v1

Each program receives a different combination of states from its bank. Shared states within a bank provide controlled comparisons between programs; different banks and data partitions have separate source episodes. Source states are selected for feature coverage and long-horizon viability. Each selected state appears under at least two programs. IDs are globally unique strings, not integers.

Random programs use uniform action choices held for 4–16 frames. Policy programs are frozen tapes from a ball-following heuristic with 20% exploratory decisions, held for 4 frames. The policy is not an expert or learned agent. Both retain every intermediate frame: ALE frame skip is 1 and sticky-action probability is 0.

Short lengths are balanced across 256/320/384/448/512; long lengths across 576/640/704/768. Length is fixed before candidate rejection. Accepted trajectories are never reset, padded, or shortened. Life loss is allowed; game-over is excluded.

Consequence diversity

At steps 64, 128, 192, and 256, signatures record relative ball displacement (quantized in units of 4), changed brick bits, cumulative reward, and life changes. They exclude absolute initial state, paddle position, and clocks.

• Different retained programs on the same observed start have distinct signatures.
• Each retained program has distinct signatures across all its selected starts.
• These are acceptance criteria: 100% distinction is a property selected into the release, not an estimate for arbitrary Breakout programs or states.
• Comparisons apply only to programs actually sharing a state. Separate banks are not a fully crossed experiment over every program and every initial state.
• Distinct trajectory signatures do not guarantee distinct final endpoints, nor identification of an arbitrary program from three demonstrations.

Download and load

Install huggingface_hub and numpy. No login is needed. Pin revision to this repository's release commit for reproducibility; main follows future revisions. The release uses ZIP/NPZ rather than Parquet, so load it with NumPy:

from pathlib import Path
import hashlib, zipfile
import numpy as np
from huggingface_hub import snapshot_download

root = Path(snapshot_download("odats/breakout-ram4-diverse-v2", repo_type="dataset", token=False))
for line in (root / "SHA256SUMS").read_text().splitlines():
    expected, name = line.split("  ", 1)
    assert hashlib.sha256((root / name).read_bytes()).hexdigest() == expected

# Load one program directly from its archive; no extraction required.
import io
with zipfile.ZipFile(root / "train.zip") as archive:
    with np.load(io.BytesIO(archive.read("train/program_0000.npz")), allow_pickle=False) as f:
        data = {key: f[key] for key in f.files}
assert data["states"].shape[1] == len(data["actions"]) + 1
segment_actions = data["actions"][10:42]       # (32,)
segment_states = data["states"][0, 10:43]     # (33, 4, 128)

Arrays

For R=16 (train) or R=4 (test), and program length L:

Key	Shape	Meaning
actions	[L]	Shared tape: 0 NOOP, 1 FIRE, 2 RIGHT, 3 LEFT
ram	[R,L+1,128]	Raw uint8 RAM frames
states	[R,L+1,4,128]	Causal four-frame histories, oldest first
ram_in, ram_out, initial_history	[R,4,128]	Initial/final histories
rewards, terminal	[R,L]	Rewards and game-over flags
lives	[R,L+1]	Remaining lives
start_ids, held_out	[R]	Globally unique source IDs and validation/test flag
warmup_actions	[3]	FIRE actions preceding the program
warmup_rewards, warmup_lives	[R,3], [R,4]	Warm-up observations
policy_source_initial	variable bytes	Full emulator snapshot for policy-tape source
policy_source_ram, policy_source_rewards	[L+1,128], [L]	Only in policy splits

Verification and observed events

An independent audit reloads all data, reconstructs source episodes, and exactly replays every trajectory from serialized emulator states. It recomputes signatures, checksums, histories, and event gates. Negative controls detect copied-program and state-invariant effects. Full machine-state reproducibility does not prove that RAM histories contain all Markov state.

Split	Reward-positive replays	Life-loss replays	Brick-change replays	Minimum moving fraction
train	74.9%	96.1%	65.6%	50.0%
random_matched	74.1%	95.1%	63.5%	50.2%
random_long	97.4%	99.8%	89.5%	52.8%
policy_matched	78.4%	95.0%	72.0%	62.9%
policy_long	98.7%	99.8%	92.9%	62.8%

Exact replay checks: 20,480; policy source checks: 512. Initial RAM/history and source-episode overlap across partitions: zero.

Observation aliasing: 2,570 of 7,856,711 distinct observed four-frame-history/action keys have multiple recorded next histories. Thus four RAM frames plus the next action are not sufficient to make every observed transition deterministic in this release. Full snapshots remain exactly reproducible.

Including the previous three actions leaves 269 observed conflicting keys out of 7,951,381 distinct extended keys. This is an empirical check, not a guarantee for unseen states.

The complete audit.json reports cross-partition intermediate RAM/history and endpoint overlap, exact action-tape substring containment, and observed history/action transition ambiguity. These are measurements, not assumptions of complete state-space separation. Approximate near-duplicates are not exhaustively excluded.

Research use and limitations

Use the held_out flag for train/validation assignment; keep all four test splits out of training and hyperparameter selection. When creating substring examples, assign partitions first and keep every window from its source trajectory in that partition. Report the number of independent banks as well as programs and replays. Use several optimization seeds and bank-level resampling for uncertainty estimates. The test splits jointly change programs and source states; validation checks known programs on held-out starts. These tests do not isolate every generalization axis. Compare ERD and action-conditioned models with their different input information clearly stated; byte accuracy alone can be dominated by unchanged RAM. Include copy-input, changed-byte, ball-motion and exact-output metrics.

The initial-state distribution is deliberately selected by fixed viability probes, feature coverage, active-ball/life constraints, and consequence diversity. The source policy, survival rule, and event requirements bias the population. This benchmark supports controlled transformation experiments; it is not an unbiased sample of game play, a standard sticky-action control benchmark, or evidence of general gameplay competence. See PROTOCOL.md for exact rules and all pre-generation pilot revisions. PROTOCOL_ADDENDUM.md and global_state_deduplication.json document deterministic replacement of banks containing duplicate initial states; all replacement banks are independently re-audited.

The generation seed is 20260917. ALE 0.12.1 was used; the manifest records the ROM hash and source-code hashes. Generate with python generation/generate.py --out NEW_DIR, then audit with python generation/audit.py NEW_DIR. Install generation/requirements.txt first. banks.zip stores per-bank selection and rejection records; snapshots.zip contains pre/post-warm-up emulator snapshots. No Atari ROM is included, and no license is asserted for third-party game content or emulator-state contents. A compatible ROM is needed only to regenerate or replay, not to load the NumPy trajectories.

References: ALE, evaluation protocols.