declref-01-declref_01_symbols-2B

Procedurally generated decl-ref-01 documents — the scoped declare/reference language of declref, with noisy references, sparse part-transition chains, interleaved openings, and periodic topic shifts tuned so that a model trained on it matches natural-language / code entropy dynamics (positional entropy profile, its fluctuation texture, and the entropy-quantile distribution), and holds that match as training doubles. Each document is a random program of { } scopes, DECL name value statements and REF name value statements whose value tokens repeat an earlier declaration — predictable only by attending back to it. Names and values are 1–3 subtoken symbols. Token ids: 0=OPEN 1=CLOSE 2=DECL 3=REF, then name-parts and value-parts in disjoint blocks; vocab = 1,028.

Grammar parameters

param	value
seq_length (document)	2048
shards (canonical layout)	256

file	split	tokens
train.bin	train	2,000,683,008
val.bin	val	9,961,472

The bin is a concatenation of fixed-length 2048-token documents (num_tokens is a whole multiple of it): read via tokens.reshape(-1, 2048) to train one document per row. Document alignment matters — a window starting mid-document orphans its references. train (seed 0) and val (a disjoint seed) are independent streams of the same grammar; token count = filesize / 2; the metas carry the full config, including the shard layout that fixes the bin's exact content.

Load a bin with the standard Hugging Face downloader:

from huggingface_hub import hf_hub_download
import numpy as np

path = hf_hub_download(repo_id="alexkstern/declref-01-declref_01_symbols-2B", filename="train.bin", repo_type="dataset")
tokens = np.memmap(path, dtype="uint16", mode="r")