blab-jhu/KYS-Pre-Rewritten

Know Your Sources: raw baseline corpora (raw text)

Seven raw, unrewritten training corpora for the Know-Your-Sources 1.5B grid. Each is about 10B training tokens. They come in two families, documented in reports/GLOBAL_TOP10B_SELECTION_REPORT.md (global Top-10B) and reports/RAW_SELECTED_BASELINES_PROVENANCE.md (strategy-linked).

Strategy-linked raw controls (4)

Composition: the shared 5B anchor (4,120,164 documents / 5,000,002,332 training tokens, identical in all four and to the anchor inside each rewritten arm) plus a 5B raw strategy half: a seed-42 uniform random sample, whole documents, of the unique source documents of one rewritten arm's final rewritten half in wytro/Know-Your-Sources@9e5ff241, kept as original text.

These are equal-token-budget controls, not identical-document controls. Rewriting roughly halves document length, so at the same 5B budget the raw half holds a random ~half of the rewritten half's sources (same distribution, not the same documents). Only sources with a successful rewrite are eligible.

folder	rewritten counterpart	raw strategy half	comparison it supports
raw_text/raw_diversity_oriented/	diversity_oriented	random 51.4% of the rewritten half's sources	rewrites vs original text of this strategy's sources, equal tokens
raw_text/raw_disagreement_aware/	disagreement_aware	random 51.8%	same
raw_text/raw_random/	wrap_inspired	random 52.4% of WRAP's input documents, restricted to those WRAP rewrote successfully (99.7% of the input)	WRAP-style rewriting of a uniform sample vs a uniform raw sample of the same population (pool minus validation and anchor); not WRAP's exact documents
raw_text/raw_rewire_inspired/	rewire_inspired	random 45.9% of the sources whose rewrites passed REWIRE's post-rewrite fastText filter	conditional: rewritten vs original text given the filter's picks

raw_rewire_inspired is a conditional ablation. Its membership depends on how each document's rewrite scored, which exists only after rewriting. It does not support "REWIRE beats raw data" and is not a raw-only selection policy; for REWIRE's pipeline effect compare rewire_inspired with raw_random (same input population).

Quality-First has no raw arm. Its raw comparison is the existing Quality-Base: the anchor + the fastText-best 5B of Quality-First's own rewriting input.

Global Top-10B controls (3, no anchor)

The whole ~10B corpus is one global Top-10B selection over the same universe as the original 1.5B Quality-Base: the 99,949,162 scored documents of blab-jhu/KYS-DCLM-Refinedweb-100M-Scored, minus the 50,000-doc validation holdout.

Selection rule:

• score descending, with the original seeded tie-break;
• whole documents, until cumulative training tokens first reach 10B.

What these corpora do not have: no shared anchor, no rewriting, no floors, quotas, variance terms or domain restrictions.

Comparator: the existing fastText Quality-Base. Rebuilt under the same conventions, its document set reproduces the published Quality-Base digest, and it equals the global fastText Top-10B plus 3 tail documents.

Scorer-training data. The ModernBERT quality head was fit on 50,427 Claude-labelled documents; all DCLM ones were removed from the scored pool before scoring, so none is in any corpus. A separate ~5M-document analysis sample was not used to fit the head and was not removed, exactly as in the original Quality-Base universe; every selection holds it at its pool rate (5.00%). Benchmark contamination was not tested. Details: selection report §2b.

folder	score
raw_text/raw_top10b_fineweb_edu/	fineweb-edu-ranking-v2 (tie-aware global percentile)
raw_text/raw_top10b_modernbert/	modernbert-ranking-v2
raw_text/raw_top10b_consensus/	mean of the three percentiles: (fastText + FineWeb-Edu + ModernBERT) / 3

Files

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

column	type	meaning
orig_doc_id	int64	position in the 100M DCLM-RefinedWeb reservoir sample (shard id // 500000, row id % 500000)
doc_id	int64	global Top-10B only: row of the scored pool, the selection key
source	string	anchor / strategy (strategy-linked) or selected (global Top-10B)
text	string	the original document text

• Each folder is the final training corpus, in training order. Documents are shuffled once at the document level (seed 42, pp_io.bucketed_shuffle), and rows follow that order file by file.
• manifest.json records every count, each setting's expected_total_tokens, the sha256 of every file, the tokenizer's sha256, and the generation code commits.
• settings_overview lists every setting's family, anchor presence and comparator.
• selection/<setting>/ holds the selected doc ids of the global Top-10B settings. The digest conventions are in manifest.json under global_top10b.

Tokenization — done by the consumer

Use tools/kys_raw/tokenize_raw_text.sh <data_root> <setting> from imHuicongZhang/nanotron (branch huicong-dev).

What the script does:

• runs 16 datatrove tasks; each document becomes its tokens + one </s> (id 2), with no <s> (datatrove replaces the tokenizer's default <s>-prepending post-processor);
• does no merging or shuffling;
• runs tools/fix_ds_metadata.py;
• checks the total against expected_total_tokens.

Token convention: len(llama2_tokenizer(text, add_special_tokens=False)) + 1 per document.

Check before publication: for all seven settings, the datatrove totals from this exact script equal expected_total_tokens (selection report §8; provenance report §R5). The script enforces the same check for every setting.

Procedure: configs/1.5B-baseline/WORKFLOW_RAW_BASELINES.md in the code repository.

Provenance in brief

• Text. Every text is read from the raw 100M pool by position. No rewritten text is used anywhere.
• Global Top-10B settings:
  • the selection code uses the original 1.5B selection primitives (universe, tie-break, whole-document cutoff); applied to fastText it reproduces the published Quality-Base document-set digest;
  • the three percentile columns recompute exactly from the raw scores (all 99,949,162 rows);
  • every document was re-tokenized at assembly and its length checked equal to the scored pool's tokens-llama2;
  • exported files were checked against the selection (document set, no duplicates, order, token total) and a sample of texts was compared byte for byte with the raw pool (selection report §7b).
• Code: tools/kys_raw/ in the code repository. The exact commits are in manifest.json.

Related

• Tokenizer: tokenizer/ in this repo, the exact llama-2 tokenizer directory the grid used.
• Init checkpoints: wytro/Know-Your-Sources-init.
• Rewritten comparators (v2): wytro/KYS-1.5B-Rewritten-v2.
• Trained raw baselines: blab-jhu/KYS-1.5B-Raw-Selected-Baselines.