blab-jhu/KYS-Pre-Rewritten

Know Your Sources: raw-selected baseline corpora (raw text)

No-rewrite controls for the Know-Your-Sources 1.5B grid. Each corpus trains a model on the same shared 5B anchor and the same selected source documents as a rewritten arm, but keeps those documents in their original, unrewritten form, subsampled at the document level to the same 5B-token strategy budget. Training on these against the rewritten arms separates the effect of rewriting from the effect of source selection.

folder	rewritten counterpart	strategy half
raw_text/raw_diversity_oriented/	diversity_oriented	original text of the documents the Diversity-Oriented arm rewrote
raw_text/raw_disagreement_aware/	disagreement_aware	original text of the documents the Disagreement-Aware arm rewrote
raw_text/raw_random/	wrap_inspired	original text of the uniform random sample the WRAP-Inspired arm rewrote
raw_text/raw_rewire_inspired/	rewire_inspired	original text of the source documents of the 5B kept by REWIRE's post-rewrite filter

raw_random is the no-selection reference for wrap_inspired: its sources were drawn uniformly at random with no quality or diversity criterion, so it is both the no-rewrite control for wrap_inspired and the reference for what selection adds. raw_rewire_inspired starts from a random pool too, but its sources are only the documents whose rewrites passed REWIRE's post-rewrite filter — a selection made after rewriting.

Files

Each raw_text/<setting>/ holds 16 parquet files, part-00000.parquet … part-00015.parquet:

column	type	meaning
orig_doc_id	int64	position of the document in the 100M DCLM-RefinedWeb reservoir sample
source	string	anchor (the shared 5B) or strategy (this setting's 5B)
text	string	the original document text

This is the final training corpus, in training order: anchor and strategy documents are already merged and shuffled; rows follow the exact post-shuffle document order, file by file. Each folder is about 10B tokens. manifest.json records every count, the expected token total after tokenization, and the sha256 of every file.

Tokenization is done by the consumer, with tools/kys_raw/tokenize_raw_text.sh from imHuicongZhang/nanotron (branch huicong-dev), as described in configs/1.5B-baseline/RUNBOOK.md: 16 datatrove tasks, one </s> per document, no merging or shuffling, then tools/fix_ds_metadata.py, then a check against the manifest's token total.

How the corpora were built

1. Source documents. For each rewritten arm, the orig_doc_id of every non-anchor row (source_prompt != 'original') of the published wytro/Know-Your-Sources parquet, deduplicated across the rewriting prompts: the source documents whose rewrites were kept.
2. Original text. Read from the 100M reservoir sample by position (orig_doc_id: shard id // 500000, row id % 500000). No rewritten text is used anywhere.
3. Token budget rule. Tokens per document = len(llama2_tokenizer(text, add_special_tokens=False)) + 1 (the +1 is the end-of-document token). A source set above 5,000,000,000 tokens is cut by taking documents in the order of numpy.random.default_rng(42).permutation over the sorted unique ids and keeping the shortest prefix reaching 5B — whole documents only, at most one document of overshoot. At or below 5B it is used as is. The same rule applies to all four settings.
4. Anchor. The shared 5B anchor — 4,120,164 documents, 5,000,002,332 tokens, the rows with source_prompt == 'original', identical in every arm — also read from the pool by position.
5. Shuffle. Anchor and strategy documents shuffled together at the document level with seed 42 by pp_io.bucketed_shuffle, the same function every published arm was shuffled with.
6. Leakage checks, all passed before upload: every parquet read in the build code reads text only from the raw pool; all 4,120,164 anchor texts match the pool byte for byte and total exactly 5,000,002,332 tokens; 200 sampled documents per setting match the pool and match none of the rewrites of the same source document; style features (markdown headings and lists, missing URLs/boilerplate) and document lengths sit with the raw pool, not with the rewritten arm.

Code: tools/kys_raw/ (build_raw_sources.py, assemble_raw_corpus.py, verify_raw_corpus.py, publish_raw_text.py); the exact commit is in manifest.json.

Related

• Tokenizer: tokenizer/ in this repo, the exact llama-2 tokenizer directory the grid was tokenized with (sha256 of each file in manifest.json, tokenizer.sha256). tokenize_raw_text.sh uses <data_root>/tokenizer by default.
• Init checkpoints: wytro/Know-Your-Sources-init, _init_1.5B_seed{42,43,44}/0/, with hash manifests.
• Training configs: configs/1.5B-baseline-seed{42,43,44}/ and configs/1.5B-baseline/RUNBOOK.md in the code repo.