passkey_swar — Passkey + Sliding-Window Associative Recall

A benchmark for fixed-size-state sequence models (linear attention / SSMs). One sequence contains (i) a passkey that must be retained indefinitely and (ii) a stream of associative-recall (AR) writes and queries that must be forgotten (queries on keys not written for more than 2w writes have answer NULL). A single-head model with a scalar decay α cannot satisfy both: forgetting needs α < 1, indefinite retention needs α = 1.

Format

BOS  PK p1 p2 p3 p4   <stream>   PKQ p1 p2 p3 p4 EOS

• passkey: 4 tokens from v0..v15 (65,536 possibilities).
• stream: N AR writes W k v, interleaved with queries Q k a (a = answer). Queries appear at random positions: after every write, further queries follow with probability q/(1+q), q = 0.25 (≈ 1 query per 4 writes).
• keys are drawn from a drifting active set of 6–8 of 16 keys; after each write, with probability 0.1 one active key is swapped for an inactive one.

Answer rule

Let age = number of writes since the key's most recent write (the immediately preceding write has age 1; overwrites count, only the latest write matters).

key state	query?	answer	answer_type
age ≤ w = 16	yes	value of that latest write	ar_valid
age > 2w = 32, or never written	yes	NULL	ar_expired
16 < age ≤ 32	never queried (gray zone)	–	–

Expired queries are what forces decay < 1: without them a delta-rule model could hold all keys and the passkey with α = 1. Expired queries are either *never-written* keys or *stale* keys (written earlier, then idle > 2w writes); only the latter test forgetting. With N ≤ 2w = 32 no key can be stale, so n32 has only never-written expired queries (see table).

Vocabulary (vocab.json, 40 tokens)

PAD=0 BOS=1 EOS=2 PK=3 PKQ=4 W=5 Q=6 NULL=7, keys k0..k15 = ids 8..23, values v0..v15 = ids 24..39.

Configs and splits

Config n{N}, N = number of AR writes. train: n32–n256 (100,000 examples each); validation and test: all N ∈ {32, 64, 128, 256, 512, 1024, 2048, 4096, 8192, 16384} (1,000 examples each). Splits use different seeds and are checked to share no example.

Columns

• input_ids: list[int16] — the full sequence.
• labels: list[int16] — -100 except the answer token a of each Q k a and the 4 passkey tokens after PKQ. HF convention: labels[t] is the target for the token at position t (shift inside the loss).
• answer_type: list[str] — one entry per supervised position, in order: ar_valid / ar_expired per query, then passkey ×4.
• n_writes: int, seed: int64 — per-example seed = split_seed * 10**12 + N * 10**6 + index; make_example(seed, N, params) reproduces the example exactly.

Statistics (measured on the released files)

config	split	examples	mean length	queries/ex	valid	expired	stale share of expired
n32	train	100,000	132	8.0	0.500	0.500	0.000
n64	train	100,000	252	16.0	0.499	0.501	0.077
n128	train	100,000	492	32.0	0.500	0.500	0.278
n256	train	100,000	972	64.0	0.500	0.500	0.535
n32	validation	1,000	132	8.0	0.503	0.497	0.000
n64	validation	1,000	252	16.0	0.495	0.505	0.077
n128	validation	1,000	492	32.0	0.499	0.501	0.285
n256	validation	1,000	972	64.1	0.503	0.497	0.536
n512	validation	1,000	1933	128.3	0.501	0.499	0.744
n1024	validation	1,000	3850	255.5	0.498	0.502	0.873
n2048	validation	1,000	7688	510.7	0.501	0.499	0.937
n4096	validation	1,000	15371	1023.6	0.500	0.500	0.969
n8192	validation	1,000	30729	2046.9	0.500	0.500	0.984
n16384	validation	1,000	61447	4094.4	0.500	0.500	0.992
n32	test	1,000	132	8.1	0.499	0.501	0.000
n64	test	1,000	251	15.8	0.502	0.498	0.078
n128	test	1,000	491	31.8	0.498	0.502	0.281
n256	test	1,000	971	63.8	0.503	0.497	0.540
n512	test	1,000	1930	127.3	0.500	0.500	0.747
n1024	test	1,000	3853	256.2	0.500	0.500	0.874
n2048	test	1,000	7697	513.7	0.499	0.501	0.937
n4096	test	1,000	15373	1024.2	0.500	0.500	0.968
n8192	test	1,000	30726	2046.0	0.500	0.500	0.984
n16384	test	1,000	61442	4092.6	0.500	0.500	0.992

Full tables (query-gap histogram, age histograms, key-reuse statistics) are in stats.md in this repo.

Sanity checks (all passed before upload)

A brute-force oracle that re-derives every label from the token stream alone reproduces 100% of labels and answer_type on every example (30,042,291 supervised positions); it also rejects deliberately corrupted examples; gray-zone queries never occur; regenerating from the stored seed reproduces stored examples; no example is shared between splits.

Generation

Generator: make_passkey_swar.py (in this repo), git commit 8e047fc6b765d70d2eda27b3d2e429d45e6d45f3. Parameters: {"w": 16, "n_keys": 16, "n_values": 16, "query_rate": 0.25, "valid_frac": 0.5, "drift_rate": 0.1, "active_min": 6, "active_max": 8}; split seeds {"train": 1, "validation": 2, "test": 3}; versions {"python": "3.11.11", "numpy": "1.26.4", "pyarrow": "19.0.1"}.

python data/make_passkey_swar.py --w 16 --n-keys 16 --n-values 16 --query-rate 0.25 \
    --drift-rate 0.1 --seed-train 1 --seed-validation 2 --seed-test 3

from datasets import load_dataset
ds = load_dataset("irodkin/passkey_swar", "n256", split="train")