EleutherAI/fineweb-heldout-queries-2048

FineWeb held-out query set (1000 x 2048 tokens, Qwen2.5 tokenizer) (Warning: documentation is by Claude)

Held-out evaluation sequences for training data attribution on Qwen-family pretraining runs over FineWeb. Drawn from a permanently reserved pool of FineWeb documents that no training set at any scale may contain, then decontaminated against FineWeb sample-100BT. The companion Pile set is EleutherAI/pile-heldout-queries-2048.

Reserved pool (corpus-wide rule)

A FineWeb document is reserved iff splitmix64(fnv1a64(id) XOR 2026) % 1000 == 0, where id is the document's id field (its urn:uuid, identical in every dump and sample) as UTF-8, fnv1a64 is 64-bit FNV-1a and splitmix64 the SplitMix64 finaliser. The rule reserves 1/1000 of documents anywhere in FineWeb, so any training pool of any size (a sample, a dump, or all of FineWeb) must drop exactly the documents it selects. Within sample/100BT it reserves 147,948 of 147,639,585 documents (reserved_ids_sample100BT.txt).

Windows

Packed like pretraining samples: a random reserved start document (random offset if longer than 2048), then following reserved documents joined by <|endoftext|> (id 151643) up to 2048 tokens of the Qwen2.5-1.5B tokenizer. Columns: input_ids, length, text, doc_ids (source ids), dumps, n_docs, pool_coverage, pool_longest.

Quality filters

Control characters <= 0.01, unique-token ratio >= 0.3, English stopword ratio >= 0.05, alphabetic fraction >= 0.6.

Query deduplication (decontamination) against the training pool

Reservation vs. deduplication. The reservation rule above removes the windows' source documents from every training set. Deduplication additionally rejects windows whose text also appears, in whole or in large part, in other documents of the pool (FineWeb is MinHash-deduplicated per dump only), so that no training set can contain a copy of a query.

Detection unit. A token-level 13-gram exact-match scan (Qwen2.5 tokenizer, rolling polynomial hash verified against the exact candidate 13-gram set) over the 101,065,736,052 tokens of the 147,491,637 non-reserved documents of sample/100BT. For every candidate window this yields

• pool_coverage: the fraction of the window's tokens inside at least one 13-gram also present in some non-reserved document, and
• pool_longest: the longest exact shared substring in tokens (a run of r consecutive shared 13-grams is a shared substring of r + 12 tokens).

Rejection rule (Lee et al. 2022 ExactSubstr thresholds). A window is rejected if pool_longest >= 50 or pool_coverage > 0.20. The same rule is applied between candidates in order (earlier candidate wins), so the kept windows are mutually non-duplicate.

What the rule does and does not guarantee. Shared substrings shorter than 13 tokens are below the detection unit and are neither measured nor penalised; shared substrings of 13-49 tokens are tolerated as long as they cover at most 20% of the window. Sub-13-token overlaps (names, phrases, boilerplate fragments) occur in essentially every web-text window.

Outcome. Of 10,000 candidates, 9,874 passed the quality filters; of those 2,520 passed this rule and the between-window check, and 1000 were drawn uniformly (seed 2026). Residual overlap of the selected 1000 windows against the non-reserved pool:

statistic	value
longest shared substring, median / p90 / max	25 / 44 / 49 tokens
coverage by shared 13-grams, median / p90 / max	4.4% / 12.3% / 19.8%
windows sharing no 13-gram with any non-reserved document	100

Per-window values are the pool_coverage and pool_longest columns, so a stricter subset can be selected without rescanning; among the 9,874 quality-passing candidates the survivors under stricter rules are:

rule (longest < L and coverage <= C)	candidates surviving
L = 50, C = 20% (this release)	2,739
L = 32, C = 10%	1,658
L = 25, C = 5%	1,094
L = 13, C = 0% (no shared 13-gram at all)	211

decontamination_summary.json lists every rejected candidate with its reason.

Scaling to larger training pools

The reservation rule already excludes the windows' source documents from any FineWeb pool. To keep near-duplicates out of a pool larger than sample-100BT, run filter_training_pool.py over the new pool: it drops every training document that shares a >= 50-token substring with a query or covers more than 20% of a query's tokens with shared 13-grams (the same rule, applied on the training side), so the query set never changes.