Search is not available for this dataset
keys list | sequences list | query uint16 0 1.02k |
|---|---|---|
[
1396,
944,
433,
1671,
255,
272,
818,
1629,
1893,
1879,
1270,
709,
317,
1668,
852,
1385,
1824,
1359,
350,
1298,
181,
450,
1087,
1294,
1291,
171,
1641,
140,
2029,
397,
1617,
1362,
57,
1308,
1716,
1310,
1945,
1481,
75,
1867,
435,
240,
... | [
[
1,
5,
0,
6,
7,
6,
6,
11
],
[
1,
2,
10,
2,
15,
12,
3,
12
],
[
15,
6,
14,
1,
9,
10,
2,
14
],
[
12,
9,
4,
13,
7,
14,
6,
2
],
[
13,
11,
4,
14,
9,
2,
... | 1,019 |
[
335,
962,
776,
1580,
1075,
515,
224,
1529,
1573,
1547,
191,
1499,
910,
415,
1147,
138,
448,
1351,
1337,
1258,
1519,
634,
198,
841,
1031,
88,
441,
30,
1764,
1318,
1426,
1546,
86,
928,
1531,
1080,
1148,
278,
1270,
303,
704,
624,
573,
... | [
[
11,
4,
13,
4,
3,
2,
0,
2
],
[
3,
15,
11,
11,
2,
5,
0,
5
],
[
9,
11,
1,
15,
9,
9,
15,
8
],
[
10,
9,
14,
0,
10,
11,
0,
8
],
[
8,
15,
12,
11,
11,
15,... | 418 |
[
1508,
1572,
1692,
211,
1006,
872,
1318,
1053,
584,
1569,
458,
877,
264,
1878,
648,
289,
299,
783,
1804,
1351,
133,
1948,
1341,
1206,
1679,
677,
992,
1556,
522,
181,
933,
1382,
480,
1100,
1322,
1375,
1050,
1110,
611,
1344,
363,
163,
... | [
[
1,
9,
2,
8,
0,
10,
5,
4
],
[
3,
8,
15,
1,
14,
6,
5,
9
],
[
2,
12,
4,
3,
11,
1,
14,
13
],
[
12,
10,
5,
6,
9,
10,
8,
12
],
[
6,
4,
3,
10,
14,
6,
... | 337 |
[
1525,
2009,
1887,
1292,
195,
1550,
1153,
505,
1366,
776,
552,
180,
891,
804,
1967,
698,
895,
1385,
1922,
1469,
1813,
587,
124,
1279,
280,
1781,
372,
147,
1595,
1544,
1699,
467,
88,
836,
563,
351,
190,
361,
329,
671,
1597,
221,
1805,... | [
[
7,
10,
10,
5,
12,
13,
12,
12
],
[
0,
8,
0,
4,
1,
10,
15,
12
],
[
2,
5,
5,
10,
5,
9,
8,
6
],
[
3,
4,
7,
7,
6,
0,
7,
8
],
[
9,
3,
3,
8,
15,
8,
4... | 1,007 |
[
346,
1745,
308,
1273,
39,
557,
1377,
1195,
2002,
95,
1642,
1465,
191,
1977,
1733,
1170,
886,
1670,
810,
1542,
181,
991,
915,
1347,
1796,
1807,
1738,
1952,
1035,
1173,
122,
374,
1576,
1842,
251,
153,
71,
525,
720,
417,
1792,
1790,
39... | [
[
15,
4,
2,
11,
2,
0,
15,
5
],
[
3,
11,
0,
3,
14,
13,
5,
11
],
[
15,
0,
7,
0,
1,
2,
9,
6
],
[
1,
12,
13,
1,
7,
15,
13,
6
],
[
6,
4,
7,
12,
2,
2,
... | 525 |
[
832,
421,
517,
1768,
1828,
560,
1024,
1399,
187,
1914,
1219,
865,
1877,
287,
1459,
542,
1805,
1301,
495,
1567,
1334,
1198,
330,
1495,
271,
1426,
1765,
2040,
228,
900,
326,
1181,
1506,
1317,
217,
846,
3,
1631,
1625,
669,
1452,
768,
8... | [
[
7,
10,
14,
7,
12,
5,
10,
0
],
[
1,
1,
15,
14,
2,
8,
13,
15
],
[
4,
11,
5,
0,
0,
5,
0,
2
],
[
13,
3,
7,
12,
5,
4,
0,
11
],
[
0,
8,
1,
0,
13,
2,
... | 587 |
[
1374,
1180,
890,
1741,
1229,
153,
1817,
527,
1399,
1713,
470,
124,
442,
960,
349,
1337,
1954,
836,
1848,
1183,
980,
1426,
913,
116,
301,
474,
684,
277,
1336,
1683,
1851,
1828,
1468,
1581,
1993,
73,
520,
345,
9,
1698,
532,
1519,
663,... | [
[
11,
10,
7,
15,
0,
11,
6,
14
],
[
5,
12,
1,
11,
7,
2,
0,
8
],
[
6,
7,
7,
4,
10,
14,
10,
2
],
[
1,
9,
0,
9,
5,
3,
0,
12
],
[
5,
6,
10,
11,
0,
5,
... | 602 |
[365,1992,1297,1146,903,897,157,1055,1773,983,998,764,738,1449,920,312,1013,509,338,1745,296,278,198(...TRUNCATED) | [[3,8,2,11,1,9,8,10],[8,12,5,10,0,4,9,14],[9,8,15,15,13,13,10,14],[6,1,0,15,14,3,7,6],[4,5,14,5,1,6,(...TRUNCATED) | 85 |
[2037,1171,1065,1965,751,1726,1620,1569,1048,1787,1085,1560,726,524,755,1009,367,2021,1373,1151,865,(...TRUNCATED) | [[1,4,13,7,6,13,3,3],[12,5,3,11,14,5,10,8],[9,2,0,15,4,1,11,7],[6,7,7,15,10,12,4,2],[2,11,9,9,12,15,(...TRUNCATED) | 634 |
[781,436,613,36,848,1102,142,1376,1494,1201,2013,649,1372,1878,695,1850,316,943,1609,1442,345,1942,4(...TRUNCATED) | [[0,4,12,0,3,1,1,14],[2,10,14,15,13,9,8,0],[1,4,12,1,6,15,12,14],[12,10,12,4,4,9,5,2],[3,4,3,4,1,0,2(...TRUNCATED) | 241 |
End of preview. Expand in Data Studio
Keyed Recognition Memory
A benchmark for comparing a memory that recalls with one that recognises, as the number of stored items grows. Each sample stores N sequences of length 8 under N keys, N = 1, 2, 4, …, 1,024. A query names one key and gives a candidate: recall means producing that key's sequence; recognition means telling whether the candidate is it. It is the keyed version of irodkin/recognition-memory, which stores one sequence.
Subsets
N1_L8_V16, N2_L8_V16, N4_L8_V16, N8_L8_V16, N16_L8_V16, N32_L8_V16, N64_L8_V16, N128_L8_V16, N256_L8_V16, N512_L8_V16, N1024_L8_V16: N sequences of length 8 over 16 tokens, N = 1, 2, 4, …, 1,024.
Splits: train 100,000 / validation 5,000 / test 10,000 samples.
Fields
| field | type | description |
|---|---|---|
keys |
uint16[N] |
N distinct key ids, drawn without replacement from 0..2047; their order is the storage order |
sequences |
uint8[N][8] |
the sequence stored under each key; tokens i.i.d. uniform over 0..15, the N sequences of a sample distinct |
query |
uint16 |
index (0..N-1) of the queried pair: the query key is keys[query], the true candidate sequences[query] |
candidates |
uint8[8][8] |
validation and test only: 4 small edits of the true sequence, then 4 distant negatives (below) |
num_changed |
uint16[8] |
validation and test only: positions where each candidate differs from the true sequence |
The train split has no candidates: draw training negatives with the same process (below), fresh each time.
Negatives
- Small edits (4 per query). One position chosen uniformly at random is always changed, and each of the other 7 positions with probability 1/7, so about 2 tokens change. A changed token gets a uniformly random different value.
- Other keys (min(4, N−1) per query). The sequences stored under other keys of the same sample, chosen at random. They test whether the memory binds a sequence to its key; a memory that only knows which sequences it stored can't reject them.
- Fresh (4 − min(4, N−1) per query, so only for N < 5). Uniformly random sequences that the sample does not store. With N = 1 all 4 distant negatives are fresh, with N = 2 three are, with N = 4 one is.
Evaluation protocol
- Give the memory the N (key, sequence) pairs, then the query key.
- Score the true sequence and each of its 8 candidates.
- Report on the test split:
- acc9: the share of samples where the true sequence scores strictly highest among the 9. Ties for the top score count as 1/(number tied), so a constant score gives 1/9.
- pairwise, separately for edits, other keys and fresh sequences: the share of (true, negative) pairs ranked correctly, ties counting 1/2.
- For recall: the share of samples where greedy decoding reproduces the whole sequence.
Reproduction
Generated with generate_keyed.py (base seed 20261108) and checked with test_keyed.py; every chunk of 10k samples is
seeded from (seed, N, length, vocabulary, key pool, split, chunk) and can be regenerated on its own.
- Downloads last month
- -