odats/breakout-ram4-segments-v3

Breakout RAM4 v3: segment-pool train tapes, v2 test splits

Deterministic Breakout (ALE, frame skip 1, no sticky actions) "programs" = frozen action tapes, each replayed from several selected initial states, saved as 4-frame RAM histories. Built for ERD-style analogical learning (infer a program from (before, after) demonstrations on other starts, apply it to a new state) and its substring / exact dedup training: any contiguous substring of a tape is a program, and identical action substrings across tapes are the same program seen from more starts.

What is new versus odats/breakout-ram4-diverse-v2: the 1024 train tapes are concatenations of 3–4 segments from a fixed pool of 256 random-hold segments (64 each of 64/128/192/256 actions), so every segment recurs in ~14 tapes and, with 16 starts per tape, each segment and each of its substrings is observed from ~224 starts instead of 16. Whole tapes are distinct. The four test splits are byte-identical copies of v2 (revision 61477a33175eb70cbf9fac4aa33eb9e76df3cf40), so v2- and v3-trained models are directly comparable on them; v3 train starts are disjoint from every v2 start (train, validation and test). See PROTOCOL_SEGMENTS.md.

Split	Programs	Replays	Initial states	Lengths	Tapes
train	1024	16384	1152	256–512	segment-pool splice (new)
random_matched	256	1024	128	256–512	v2 copy
random_long	256	1024	128	576–768	v2 copy
policy_matched	256	1024	128	256–512	v2 copy
policy_long	256	1024	128	576–768	v2 copy

Train: 32 independent banks of 32 fitting + 4 validation starts; each program replayed on 15 fitting + 1 validation start (held_out flag). Segment reuse: 0–46 tapes per segment (mean 13.9), 2445 distinct ordered segment pairs, parts per tape {4: 493, 3: 531}.

Split	Positive reward	Life loss	Motion minimum	Copy-input endpoint byte accuracy
train	77.2%	95.6%	50.0%	87.79%
random_matched	74.1%	95.1%	50.2%	87.92%
random_long	97.4%	99.8%	52.8%	86.13%
policy_matched	78.4%	95.0%	62.9%	87.77%
policy_long	98.7%	99.8%	62.8%	86.23%

Verification (independent audit, audit.json)

• 20,480 exact trajectory replays from serialized emulator snapshots, every RAM byte, reward, life count and terminal flag compared; 512 policy-source replays and 4,992 state-qualification probe replays passed.
• All retained programs have distinct 256-frame relative-consequence signatures across their starts, and different programs sharing a start have different signatures (explicit acceptance gates, as in v2).
• Initial RAM/history overlap between partitions: 111968; endpoint-pair overlap: 0; shared intermediate RAM frames are listed per partition pair in STATISTICS.md.
• Exact cross-split substring containment (test tape inside a train tape or vice versa, >= 256 actions): 0 cases.
• Three-demo endpoint ambiguity on train: 0 conflicting common queries.
• Train rejections during generation: {"terminal_replay": 901, "same_state_effect_collision": 3237, "low_motion": 130, "insufficient_qualifying_starts": 18, "required_underused_state_missing": 16, "across_state_effect_collision": 30}.

Caveats

• Train tapes share segments by design: programs are not independent draws. Report uncertainty over banks and over segments, not over programs.
• Test tapes are independent random / policy tapes (v2). Training on recurring segments may or may not transfer to them; that is the experiment this release supports.
• Selected for viability, coverage and responsiveness; not natural gameplay and not the sticky-action Atari setting.
• Four RAM frames plus an action are not always a sufficient state (see STATISTICS.md, observation ambiguity).

Files

train.zip (new), random_matched.zip, random_long.zip, policy_matched.zip, policy_long.zip (v2 copies): <split>/program_XXXX.npz with actions [L], states [R, L+1, 4, 128], ram_in / ram_out [R, 4, 128], ram, rewards, lives, terminal, start_ids, held_out, warm-up arrays. segment_pool.npz / segment_pool.json: the 256 segments; programs.jsonl records each train tape's segment ids (source.parts). snapshots.zip / banks.zip: emulator states and bank records for the audit. generation/: the exact code (generate_segments.py wraps the unchanged v2 generate.py).

import io, zipfile, numpy as np
from huggingface_hub import hf_hub_download
p = hf_hub_download("odats/breakout-ram4-segments-v3", "train.zip", repo_type="dataset")
with zipfile.ZipFile(p) as z, np.load(io.BytesIO(z.read("train/program_0000.npz"))) as f:
    print(f["actions"].shape, f["states"].shape)   # (L,), (16, L+1, 4, 128)