Datasets:
KYS-DCLM-Refinedweb-100M-Scored
The candidate document pool behind Know Your Sources: Data Selection Matters when Rewriting for Data-Constrained Pretraining — 99,949,162 web documents sampled from DCLM-RefinedWeb, each annotated with three independent quality scores, their tie-aware global percentiles, a 24-way WebOrganizer topic label, and a Llama-2 token count.
Every source-selection strategy in the paper is a different way of ranking this table.
Contents
200 Parquet shards, ~170 GB, 500,000 rows each (last shard 449,162).
| Column | Type | Description |
|---|---|---|
orig_doc_id |
int64 | Positional index into the original 100M reservoir sample: shard_index * 500000 + row_index. The raw pool has no native ID. |
doc_id |
int64 | Contiguous 0 … 99,949,161. This is the join key used by every selection script in the paper. |
text |
string | Document text. |
url |
string | Source URL. |
metadata |
string | Raw WARC JSON: WARC-Record-ID, WARC-Target-URI, WARC-Date, WARC-Block-Digest, WARC-Payload-Digest, WARC-IP-Address, WARC-Truncated, Content-Length, and others. |
fasttext |
float32 | DCLM fastText classifier, raw P(high-quality). |
fineweb-edu |
float32 | FineWeb-Edu classifier, raw regression logit. |
modernbert |
float32 | Our distilled ModernBERT ridge scorer, raw score. |
topic |
string | One of 24 WebOrganizer categories. |
fasttext-ranking-v2 |
float32 | Tie-aware global percentile of fasttext, in (0, 1]. Higher = better. |
fineweb-edu-ranking-v2 |
float32 | Same for fineweb-edu. |
modernbert-ranking-v2 |
float32 | Same for modernbert. |
tokens-llama2 |
int32 | len(llama2_tokenizer(text, add_special_tokens=False)). |
On the row count
The repo name says 100M because the pool was reservoir-sampled at exactly 100,000,000 documents.
This table holds 99,949,162 of them: 50,838 rows were removed because they overlap the
document set used to train the ModernBERT scorer (see
KYS-Claude-Haiku-50K-Labeled).
Removing them keeps the scorer honest on the pool it is applied to. This is the exact table every
selection strategy in the paper ran over, and the percentiles below are computed on exactly these
rows — so this, not the raw 100M, is what reproduces the paper.
On the percentiles
*-ranking-v2 are tie-aware average-rank percentiles computed over all 99,949,162 rows:
pct = scipy.stats.rankdata(score, method="average") / 99_949_162
Ties matter here: the DCLM fastText score has an 8,655,073-document tie at its floor, which maps to a single percentile value of ≈ 0.0433.
The two derived quantities the paper uses
Consensus quality q and scorer disagreement v are not stored — they are cheap to derive and
are computed on the fly by the selection code. Reproduce them exactly:
import numpy as np
ft = tbl["fasttext-ranking-v2"]; fw = tbl["fineweb-edu-ranking-v2"]; mb = tbl["modernbert-ranking-v2"]
q = ((ft + fw + mb) / 3.0).astype(np.float32) # consensus quality
v = (((ft-q)**2 + (fw-q)**2 + (mb-q)**2) / 3.0).astype(np.float32) # disagreement, ddof=0
u = q + 0.5 * np.sqrt(v) # DISAGREEMENT-AWARE utility, λ=0.5
float32 and ddof=0 are load-bearing — that is the code that built the released corpora.
Usage
from datasets import load_dataset
ds = load_dataset("blab-jhu/KYS-DCLM-Refinedweb-100M-Scored", split="train", streaming=True)
print(next(iter(ds)))
Reading a couple of shards directly is usually faster than materialising 170 GB:
import pyarrow.parquet as pq
from huggingface_hub import hf_hub_download
p = hf_hub_download("blab-jhu/KYS-DCLM-Refinedweb-100M-Scored",
"merged_clean_00000.parquet", repo_type="dataset")
tbl = pq.read_table(p)
The three scorers
Only the ModernBERT scorer is ours. Cite the other two; the pinned revisions below are the exact ones used.
| Scorer | Repo | Revision |
|---|---|---|
| DCLM fastText | mlfoundations/fasttext-oh-eli5 |
cd8b714a90f2dbcd3b02cf5fc972e5d7c7f4f107 |
| FineWeb-Edu classifier | HuggingFaceFW/fineweb-edu-classifier |
284663cbb2dabf9bda30d8f8cc49601251ee1631 |
| ModernBERT ridge (ours) | blab-jhu/KYS-Modernbert-Quality-Scorer |
— |
| WebOrganizer topics | WebOrganizer/TopicClassifier-NoURL |
8f60a8f97177f5664a81ad1c2789412fbb38351d |
The topic classifier is the NoURL variant and requires transformers 4.46.x — on 5.x its RoPE
attention silently misbehaves.
The 24 topic labels
Adult · Art & Design · Crime & Law · Education & Jobs · Entertainment · Fashion & Beauty · Finance & Business · Food & Dining · Games · Hardware · Health · History · Home & Hobbies · Industrial · Literature · Politics · Religion · Science & Tech. · Social Life · Software · Software Dev. · Sports & Fitness · Transportation · Travel
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-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}
}
- Downloads last month
- 69