KYS-Claude-Haiku-50K-Labeled

The LLM-annotated corpus used to distil the ModernBERT quality scorer in Know Your Sources: Data Selection Matters when Rewriting for Data-Constrained Pretraining.

50,427 documents, each scored by Claude Haiku 4.5 on a five-criterion rubric.

The repo name rounds to 50K. The true row count is 50,427.

Composition

Source	Documents
DCLM-RefinedWeb (mix = "dclm-rw")	49,998
OpenWebMath (mix = "openwebmath")	218
algebraic-stack (mix = "alg_stack")	211
Total	50,427

The small maths/technical addition exists so the scorer sees STEM-style text, which is rare in raw web data.

Files

File	Rows	Purpose
combined.jsonl	50,427	everything — cross-validation ran on this
train.jsonl	40,340	reporting split only (val_ratio 0.2, seed 42)
val.jsonl	10,087	reporting split only

The train/val split exists for descriptive statistics. The released scorer was fitted and evaluated by 5-fold CV over all 50,427 rows, repeated across 5 seeds.

Schema

Field	Type	Description
doc_id	string	"f<file_idx>:r<row_idx>", e.g. "f44:r270968" — position in the original 100M reservoir sample
Claude_Haiku_total_score	int	0–5, the sum of the five criteria. This is the regression target.
coherence	int	0/1 — readable, well-formed, not badly fragmented or repetitive
informativeness	int	0/1 — meaningful non-trivial content rather than boilerplate, spam or filler
internal_reliability	int	0/1 — internally consistent, not obviously fabricated (no outside knowledge used)
lm_usefulness	int	0/1 — useful for learning language or knowledge patterns
exceptional_value	int	0/1 — clear explanation, dense information, strong structure or high-quality writing
justification	string	free-text rationale, ≤ 100 words
text	string	document text, truncated to 4,096 characters for annotation
url, domain, char_length		web-document fields (dclm-rw rows)
selection_source	string	"random" or "diversified"
mix	string	"dclm-rw" / "openwebmath" / "alg_stack"

Maths rows substitute _dolma_sample_id / meta / date for the web-specific fields.

Relationship to the 100M pool

doc_id maps to orig_doc_id in KYS-DCLM-Refinedweb-100M-Scored:

f, r = doc_id[1:].split(":r")
orig_doc_id = int(f) * 500_000 + int(r)

These 49,998 web documents are exactly the rows removed from that table (50,838 rows once near-duplicates are included), so the scorer is never applied to documents it was trained on. That is why the scored pool has 99,949,162 rows rather than 100,000,000.

Annotation

Model claude-haiku-4-5-20251001 via the Anthropic Batch API, temperature 0, max_tokens=200. The system prompt scores five independent binary criteria and returns JSON:

{"coherence": 0, "informativeness": 0, "internal_reliability": 0,
 "lm_usefulness": 0, "exceptional_value": 0, "total_score": 0, "justification": "max 100 words"}

Guidelines given to the annotator: focus only on the text; do not reward length; do not penalise technical content if clear; strongly penalise navigation text, cookie banners, spam, link lists, SEO content, forum fragments, image metadata and duplicated patterns.

Unlike the single-axis educational-value scale of FineWeb-Edu, this rubric is deliberately broader — it targets general pretraining utility rather than educational content alone.

Sampling

Two equal halves of 24,999 web documents each: a uniform reservoir sample over all documents longer than 200 characters, and a length-diversified sample that partitions documents into five character-count quantile buckets and caps any single URL domain at 5% within each bucket.

What was distilled from it

KYS-Modernbert-Quality-Scorer — a ridge head on frozen ModernBERT embeddings, reaching Spearman 0.7314 against held-out annotations.

The rest of the release

Repo	What it holds
KYS-1.5B-Quality-Base	QUALITY-BASE — non-rewritten baseline
KYS-1.5B-Quality-First	QUALITY-FIRST
KYS-1.5B-Diversity-Oriented	DIVERSITY-ORIENTED
KYS-1.5B-Disagreement-Aware	DISAGREEMENT-AWARE (λ = 0.5)
KYS-1.5B-Wrap-Inspired	WRAP-INSPIRED
KYS-1.5B-Rewire-Inspired	REWIRE-INSPIRED
KYS-Modernbert-Quality-Scorer	the distilled ModernBERT ridge quality scorer
KYS-DCLM-Refinedweb-100M-Scored	the candidate pool with all scores
KYS-Claude-Haiku-50K-Labeled	the Claude Haiku annotations behind the scorer
KYS-1.5B-Pretraining-Corpora	the shared anchor + six strategy remainders
KYS-Configs	prompts, vLLM, Nanotron and eval configs + shared init weights

Citation

@misc{kys2026,
  title  = {Know Your Sources: Data Selection Matters when Rewriting for Data-Constrained Pretraining},
  author = {TODO},
  year   = {2026},
  note   = {TODO: fill in venue / arXiv id / URL}
}