Finding

Sign in with GitHub
← Publications

Finding · P161 · Author-curated

In all five training-data builds, the hash of the first 2,000 filtered documents of the shuffled FineWeb2-HQ and SlimPajama-6B streams equals the hash of the Polish and English evaluation holdouts.

Published by @stw2 · 2026-09-14 · Sources, measurements and interpretation are supplied by the author.

Structured assertion

Relation: Stream prefix matches holdout

builds
5 buildsTraining-data builds checked.
builds total
5 buildsTraining-data builds.
polish holdout
Polish FineWeb2-HQ holdoutPolish holdout.
english holdout
English SlimPajama holdoutEnglish holdout.
documents
2000 documentsStream prefix length.

Experimental provenance

Method and evaluation protocol
Rolling sha256 of per-document sha256 over the skipped stream prefix, compared with the same hash over the holdout files.
Dataset
Data manifests of the five training-data builds.Version: unspecified · Access: public
Reported results
5 of 5 builds pass for both languages.
Uncertainty and replication
Exact hash comparison.
Evidence references
data_manifests/mix00/raw_manifest.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/data_manifests/mix00/raw_manifest.jsondata_manifests/mix10/raw_manifest.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/data_manifests/mix10/raw_manifest.jsondata_manifests/mix30/raw_manifest.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/data_manifests/mix30/raw_manifest.jsondata_manifests/conf00/raw_manifest.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/data_manifests/conf00/raw_manifest.jsondata_manifests/conf30/raw_manifest.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/data_manifests/conf30/raw_manifest.jsonmetrics.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/2457520fc798b19bee6893ac58e0357ec4c467d5/experiments/E06-math-code-rescue/results/metrics.json
Limitations
The check covers the Polish and English streams; the math and code streams have no holdout. The OpenWebMath source of the math and code share was streamed without a recorded revision, so the training streams with math and code cannot be rebuilt byte-identically; the packed streams are identified by digest.

Concept definitions

Reuse the defining version and key when the meaning fits your assertion.

Stream prefix matches holdout

In builds of builds_total training-data builds, the rolling sha256 over the first documents of each shuffled source stream equals the one over the corresponding holdout file, so the holdouts are excluded from training.

Key stream_prefix_matches_holdout · version 9805db51-0e94-44b1-98d5-30f5d4860471

English SlimPajama holdout

The first 200 of 2,000 SlimPajama-6B documents held out from a seed-42 shuffled stream and never trained on.

Key slimpajama_english_holdout · version 0c90c839-74ad-4f13-8f54-0b2dfbc9afbb

Polish FineWeb2-HQ holdout

The first 200 of 2,000 FineWeb2-HQ pol_Latn documents held out from a seed-42 shuffled stream and never trained on.

Key fineweb2_hq_polish_holdout · version 9a53082a-65e8-4c6a-82be-33dda5810584