irodkin/ar_multihop

AR Multihop Associative Retrieval

irodkin/ar_multihop is a deterministic synthetic associative-retrieval dataset with multiple queries per context. Hub rows are raw structured integer arrays; model-specific structural tokens and sequence formatting are added by a training collator.

Configurations

Configurations are named N{N}-H{H}-V16 for every combination of:

• N: 8, 16, 32, 64, or 128 total key-value edges.
• H: 1, 2, 4, or 8 edges per chain; H must divide N.
• V16: one shared alphabet of exactly 16 ordinary entity symbols, numbered 0 through 15. Structural model tokens are not part of this alphabet.

Each sample contains N/H chains. The number of distinct nodes is N + N/H. Every node is represented by a fixed-width base-16 sequence. The width is the smallest positive integer L satisfying 16**L >= N + N/H. The generator finds L with integer multiplication, so exact powers of 16 are handled without floating-point rounding.

Raw schema

For a configuration with entity width L, every row has:

Field	Arrow shape	Meaning
sample_id	scalar int64	Zero-based row ID within the split
context_keys	[N, L] int8	Keys in shuffled context-pair order
context_values	[N, L] int8	Values aligned with context_keys
query_keys	[N, L] int8	Every nonterminal node, independently shuffled
targets	[N, L] int8	Terminal entity aligned with each query
hop_distances	[N] int16	Remaining edges from query to terminal

All entries in the four entity fields are raw symbols in [0, 16). Rows do not contain BOS, EOS, PAD, MASK, separators, boundaries, KV markers, query markers, model token IDs, segmentation data, or other structural symbols.

Chains and queries

The generator samples N + N/H unique integer node IDs without replacement from [0, 16**L) and only then converts them to fixed-width base 16, retaining leading zeros. It partitions these nodes into N/H disjoint acyclic chains of H+1 nodes. Consecutive nodes form the H directed edges in each chain.

Consequently, keys are unique, values are unique, internal values exactly equal the key of the following edge, chains do not share nodes, and there are no cycles, branches, or shortcut edges. All context pairs are shuffled while preserving key/value alignment.

Every one of the N nonterminal nodes is queried exactly once. Its target is the terminal node of its chain, and its hop_distance is the correct remaining number of edges. Each distance from 1 through H occurs exactly N/H times. Query order is shuffled with a deterministic random stream independent from the context-pair shuffle, while query keys, targets, and distances stay aligned.

Splits and determinism

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

Split sizes and seeds are configurable. Split name, seed, configuration, and sample index enter a versioned SHA-256 seed derivation, giving deterministic but distinct random streams.

Model-side formatting

The dataset intentionally leaves sequence construction to the collator. The provided EnergyGradMem runner maps each raw symbol into a disjoint model-entity range and inserts distinct PAD, BOS, EOS, context, KV, query, and MASK tokens. It formats KV records, segments only between complete records, and creates one MASK prediction position per target digit. Ground-truth target digits are labels only and are never inserted into query inputs.

Generation and upload

push_ar_multihop_dataset.py validates every row before Arrow writing and uses Dataset.from_generator, avoiding a million-row Python list. Every invocation first generates small local preflight versions of the selected configurations and prints decoded examples.

# Validate all configurations without full generation.
python push_ar_multihop_dataset.py --dry-run

# Materialize full datasets locally.
python push_ar_multihop_dataset.py --local-only

# Upload directly from completed generation caches.
python push_ar_multihop_dataset.py --push

# Rebuild the complete Hub configuration index from uploaded Parquet files.
python push_ar_multihop_dataset.py --sync-hub-metadata

Use --n-values, --h-values, split-size/seed options, --output-dir, and --cache-dir to select or customize generation. Uploads use the user's existing Hugging Face authentication. Existing cards are not overwritten, completed caches are removed only after a successful save/upload, and failed persistence keeps the only completed copy available for resume. Metadata synchronization indexes all complete configurations found on the Hub, including those uploaded by earlier partial invocations.