irodkin/zoology_multihop

Zoology Multihop Associative Retrieval

irodkin/zoology_multihop is a deterministic synthetic associative-retrieval dataset with multiple queries per context. It uses integer token arrays rather than natural-language text.

Configurations

Every configuration is named N{N}-H{H}-V4096 and uses one of:

• N: 8, 16, 32, 64, or 128 total key-value edges.
• H: 1, 2, 4, or 8 edges per chain.
• V: exactly 4,096 total tokens.

The complete release contains all 20 combinations. A sample has N/H node-disjoint chains. H is both the chain length and the maximum required query hop distance.

Vocabulary allocation

The complete vocabulary has exactly 4,096 IDs. IDs 0–9 are structural:

ID	Name	Meaning
0	PAD	Batch padding
1	BOS	Sequence beginning
2	EOS	Sequence end
3	CONTEXT_START	Context boundary
4	CONTEXT_END	Context boundary
5	KV_START	KV-record beginning
6	KV_SEPARATOR	Separates a key from its value
7	KV_END	KV-record end
8	QUERY_START	Query-record beginning
9	QUERY_END	Query-record end

IDs 10–4095 form one shared, unpartitioned pool of 4,086 entity tokens. Keys, values, and targets are sampled from this same pool; there are no key-only or value-only sub-vocabularies.

Integer encoding and schema

Each KV record has fixed width five:

[KV_START, key, KV_SEPARATOR, value, KV_END]

context_input_ids is one flat sequence:

[BOS, CONTEXT_START, KV record 1, ..., KV record N, CONTEXT_END, EOS]

Each query record has fixed width three and contains no target:

[QUERY_START, key, QUERY_END]

query_input_ids is a flat query collection:

[BOS, query record 1, ..., query record N, EOS]

The Arrow schema is:

Field	Type	Length	Meaning
sample_id	int64	1	Zero-based row ID within the split
context_input_ids	list<int32>	5N + 4	Flat shuffled KV context
query_input_ids	list<int32>	3N + 2	Flat independently shuffled query records
targets	list<int32>	N	Terminal entity aligned with each query
hop_distances	list<int16>	N	Required number of mapping traversals

A collator can split query_input_ids[1:-1] into width-three records and pair their middle tokens with targets and hop_distances. Targets are never part of query_input_ids, preventing teacher-forced leakage between queries.

The dataset deliberately contains no segmentation fields or metadata. Model code may segment the flat context later, provided a model boundary never splits a width-five KV record.

Chain and query generation

For each sample, the generator samples N + N/H entity IDs without replacement and divides them into N/H chains of H+1 nodes. Consecutive nodes produce the H directed KV edges in each chain. Consequently:

• every key and every value is unique;
• internal nodes occur once as a value and once as the next key;
• chains are acyclic and node-disjoint;
• there are no branches, shortcuts, duplicate mappings, or cycles.

All N KV records are shuffled before context encoding. Every nonterminal node is queried exactly once, including chain starts and intermediate nodes. Query order uses a random stream independent from context order. Each query targets its chain's terminal node. Every effective depth from 1 through H occurs exactly N/H times.

Splits and determinism

The pregenerated split defaults match the original KV-retrieval uploader:

Split	Default rows	Default seed
train	1,000,000	42
validation	5,000	43
test	10,000	44

Split sizes and seeds are CLI-configurable. Split name, split seed, configuration, and sample index are combined with a versioned SHA-256 seed derivation, giving each split a distinct deterministic random stream.

Generation

The standalone pipeline is push_zoology_multihop_dataset.py. It first creates and validates small local Arrow datasets for every selected configuration and prints decoded examples. --dry-run stops after this preflight, --local-only materializes the requested full local datasets, and --push uploads completed configurations directly from the generation cache using the user's existing Hugging Face authentication. Local-only mode removes each configuration's generation cache only after save_to_disk() succeeds. Push mode does not create a second full copy under --output-dir; it removes the generation cache only after the upload succeeds. A failed save or upload keeps the completed cache for a resumable retry.

# Validate and inspect small Arrow versions of all 20 configurations.
python push_zoology_multihop_dataset.py --dry-run

# Materialize the complete release locally without network writes.
python push_zoology_multihop_dataset.py --local-only

# Generate and upload directly, without retaining a second full local copy.
python push_zoology_multihop_dataset.py --push

Use --n-values and --h-values for a subset, and use the split-size, split-seed, --output-dir, and --cache-dir options to override generation defaults. No access token is accepted on the command line; uploads use the existing Hugging Face login.