mahan13900/Mahan-Nano-Database

Mahan Nano Database — binary pretraining data

This dataset is derived from the Mahan mixed pretraining corpus. It is public; its token IDs can be decoded back into source text. Files are added in verified commits. The dataset is complete only when COMPLETE.json is present.

Files and split

• shards/train/part-NNNNNN.bin and .idx: training documents.
• shards/validation/part-NNNNNN.bin and .idx: validation documents; a split pair is omitted if that input shard contains no documents for it.
• receipts/part-NNNNNN.json: counts and SHA-256 checksums for one input shard.
• run_config.json: pinned input/tokenizer revisions and precise split rule.
• stats.json: exact document, text-token, EOD-token, and total-token counts, by source and split, written at completion.

Validation is selected by SHA-256 of the exact UTF-8 text: the first eight bytes, interpreted big-endian, modulo 1000 equal zero. Identical texts always share a split. All occurrences, including repeated Persian Wikipedia documents, are kept. The document and token percentages are approximate, not guaranteed to be 0.1%.

Binary format

Pairs use Megatron IndexedDataset v1. .bin is a contiguous little-endian uint16 token stream; each input document is one sequence and ends with one EOD token (ID 1). Other special-token-looking strings in source text are encoded as ordinary text; padding and BOS are not inserted. No truncation, packing to 4096, deduplication, or insertion of metadata takes place here. .idx has magic MMIDIDX\0\0, version 1, dtype code 8, sequence/document index counts, int32 sequence lengths, int64 byte pointers, and int64 document indices [0,...,number_of_documents]. Empty splits have no pair.

The pairs can be read by Megatron IndexedDataset or MaxText's Grain mmap reader. MaxText mmap_npy can also read them, but additionally needs the separate document/sample/shuffle .npy indices. This job does not generate them; MaxText can build them at startup or use indices built beforehand for the training configuration. Download the files to the training environment or mount them before use; a public Hub URL is not itself a local mmap path. The training loader must form 4096-token windows and create appropriate shifted labels, shuffling, and loss/attention masks. EOD alone does not block attention across documents. Verify these loader settings for the exact framework/version used.

Tokenizer: mahan13900/Mahan-Nano-Base, subfolder tokenizer/; its pinned revision and SHA-256 are recorded in run_config.json.