---
configs:
- config_name: N1_L8_V16
  data_files:
  - split: train
    path: N1_L8_V16/train-*.parquet
  - split: validation
    path: N1_L8_V16/validation-*.parquet
  - split: test
    path: N1_L8_V16/test-*.parquet
- config_name: N2_L8_V16
  data_files:
  - split: train
    path: N2_L8_V16/train-*.parquet
  - split: validation
    path: N2_L8_V16/validation-*.parquet
  - split: test
    path: N2_L8_V16/test-*.parquet
- config_name: N4_L8_V16
  data_files:
  - split: train
    path: N4_L8_V16/train-*.parquet
  - split: validation
    path: N4_L8_V16/validation-*.parquet
  - split: test
    path: N4_L8_V16/test-*.parquet
- config_name: N8_L8_V16
  data_files:
  - split: train
    path: N8_L8_V16/train-*.parquet
  - split: validation
    path: N8_L8_V16/validation-*.parquet
  - split: test
    path: N8_L8_V16/test-*.parquet
- config_name: N16_L8_V16
  data_files:
  - split: train
    path: N16_L8_V16/train-*.parquet
  - split: validation
    path: N16_L8_V16/validation-*.parquet
  - split: test
    path: N16_L8_V16/test-*.parquet
- config_name: N32_L8_V16
  data_files:
  - split: train
    path: N32_L8_V16/train-*.parquet
  - split: validation
    path: N32_L8_V16/validation-*.parquet
  - split: test
    path: N32_L8_V16/test-*.parquet
- config_name: N64_L8_V16
  data_files:
  - split: train
    path: N64_L8_V16/train-*.parquet
  - split: validation
    path: N64_L8_V16/validation-*.parquet
  - split: test
    path: N64_L8_V16/test-*.parquet
- config_name: N128_L8_V16
  data_files:
  - split: train
    path: N128_L8_V16/train-*.parquet
  - split: validation
    path: N128_L8_V16/validation-*.parquet
  - split: test
    path: N128_L8_V16/test-*.parquet
- config_name: N256_L8_V16
  data_files:
  - split: train
    path: N256_L8_V16/train-*.parquet
  - split: validation
    path: N256_L8_V16/validation-*.parquet
  - split: test
    path: N256_L8_V16/test-*.parquet
- config_name: N512_L8_V16
  data_files:
  - split: train
    path: N512_L8_V16/train-*.parquet
  - split: validation
    path: N512_L8_V16/validation-*.parquet
  - split: test
    path: N512_L8_V16/test-*.parquet
- config_name: N1024_L8_V16
  data_files:
  - split: train
    path: N1024_L8_V16/train-*.parquet
  - split: validation
    path: N1024_L8_V16/validation-*.parquet
  - split: test
    path: N1024_L8_V16/test-*.parquet
- config_name: N8_L1_V16
  data_files:
  - split: train
    path: N8_L1_V16/train-*.parquet
  - split: validation
    path: N8_L1_V16/validation-*.parquet
  - split: test
    path: N8_L1_V16/test-*.parquet
- config_name: N8_L2_V16
  data_files:
  - split: train
    path: N8_L2_V16/train-*.parquet
  - split: validation
    path: N8_L2_V16/validation-*.parquet
  - split: test
    path: N8_L2_V16/test-*.parquet
- config_name: N8_L4_V16
  data_files:
  - split: train
    path: N8_L4_V16/train-*.parquet
  - split: validation
    path: N8_L4_V16/validation-*.parquet
  - split: test
    path: N8_L4_V16/test-*.parquet
- config_name: N8_L16_V16
  data_files:
  - split: train
    path: N8_L16_V16/train-*.parquet
  - split: validation
    path: N8_L16_V16/validation-*.parquet
  - split: test
    path: N8_L16_V16/test-*.parquet
- config_name: N8_L32_V16
  data_files:
  - split: train
    path: N8_L32_V16/train-*.parquet
  - split: validation
    path: N8_L32_V16/validation-*.parquet
  - split: test
    path: N8_L32_V16/test-*.parquet
- config_name: N8_L64_V16
  data_files:
  - split: train
    path: N8_L64_V16/train-*.parquet
  - split: validation
    path: N8_L64_V16/validation-*.parquet
  - split: test
    path: N8_L64_V16/test-*.parquet
- config_name: N8_L128_V16
  data_files:
  - split: train
    path: N8_L128_V16/train-*.parquet
  - split: validation
    path: N8_L128_V16/validation-*.parquet
  - split: test
    path: N8_L128_V16/test-*.parquet
- config_name: N8_L256_V16
  data_files:
  - split: train
    path: N8_L256_V16/train-*.parquet
  - split: validation
    path: N8_L256_V16/validation-*.parquet
  - split: test
    path: N8_L256_V16/test-*.parquet
- config_name: N8_L512_V16
  data_files:
  - split: train
    path: N8_L512_V16/train-*.parquet
  - split: validation
    path: N8_L512_V16/validation-*.parquet
  - split: test
    path: N8_L512_V16/test-*.parquet
- config_name: N8_L1024_V16
  data_files:
  - split: train
    path: N8_L1024_V16/train-*.parquet
  - split: validation
    path: N8_L1024_V16/validation-*.parquet
  - split: test
    path: N8_L1024_V16/test-*.parquet
license: mit
task_categories:
- other
pretty_name: Keyed Recognition Memory
size_categories:
- 1M<n<10M
---

# Keyed Recognition Memory

A benchmark for comparing a memory that *recalls* with one that *recognises*, as the number and the length of stored
items grow. Each sample stores N sequences of length L under N keys. 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](https://huggingface.co/datasets/irodkin/recognition-memory), which stores one sequence.

## Subsets

- **Number of items:** `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, N = 1, 2, 4, …, 1,024.
- **Length of items:** `N8_L1_V16`, `N8_L2_V16`, `N8_L4_V16`, `N8_L8_V16`, `N8_L16_V16`, `N8_L32_V16`, `N8_L64_V16`, `N8_L128_V16`, `N8_L256_V16`, `N8_L512_V16`, `N8_L1024_V16`: 8 sequences of length L, L = 1, 2, 4, …, 1,024. `N8_L8_V16` belongs to both families.

Tokens are over 16 values. 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][L]` | 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][L]` | 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 L−1
  positions with probability 1/(L−1), so about 2 tokens change (exactly 1 at L = 1). 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 (distinctness is checked for L ≤ 15; beyond that a repeat is vanishingly rare). With N = 1 all 4 distant negatives are fresh, with N = 2 three are, with N = 4 one is.

## Evaluation protocol

1. Give the memory the N (key, sequence) pairs, then the query key.
2. Score the true sequence and each of its 8 candidates.
3. 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.