Finding

Sign in with GitHub
← Publications

Finding · P192 · Author-curated

After 150M tokens of embeddings-only training, the FVT−FOCUS gap of APT4 transplants of Qwen2.5-1.5B is VANISHED on 5, SHRINKS on 2 and PERSISTS on 0 of 7 domains whose time-zero interval excludes zero.

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

Structured assertion

Relation: Counts per persistence bin

metric
Bits per byteQuantity compared.
model
APT4 transplant of Qwen2.5-1.5BModels measured.
subject
Frequency-based Vocabulary TransferInitialisation whose value is the minuend.
comparator
FOCUSInitialisation whose value is subtracted.
domains
the English SlimPajama holdout; English LaTeX method sections; English Python code; Polish Wikipedia science articles; the Polish FineWeb2-HQ holdout; Polish reviews; Polish PES examination questionsDomains classified.
training tokens
150000000 tokensTokens of embeddings-only continued pretraining, nominal checkpoint size, before the value is measured.
count total
7 domainsDomains classified.
count vanished
5 domainsDomains in VANISHED.
count shrinks
2 domainsDomains in SHRINKS.
count persists
0 domainsDomains in PERSISTS.

Experimental provenance

Method and evaluation protocol
Count the pre-registered per-domain persistence bins; the design has no class verdict for this endpoint.
Dataset
First 200 documents of the evaluation domains, scored at the 150M-token checkpoints of embeddings-only training on the packed APT4 stream.Version: unspecified · Access: restricted
Reported results
en: VANISHED; sci_latex: VANISHED; sci_python: VANISHED; pl_wiki_sci: VANISHED; pl: VANISHED; pl_informal: SHRINKS; pl_pes: SHRINKS
Uncertainty and replication
Per-domain 95% paired bootstrap intervals.
Limitations
One training run per initialisation at 1.5B. The trainer's loss and gradient-norm logs, the smoke-run gate results, the transfer digests and the token stream's digest were not kept. Windows of 2,048 tokens.

Author’s note

Values from results/analysis_s2.json, S2S_focus_persistence.<domain>.bin.

Concept definitions

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

Counts per persistence bin

Of count_total domains, count_vanished are in VANISHED, count_shrinks in SHRINKS and count_persists in PERSISTS under the pre-registered persistence rule for subject's metric minus comparator's metric after training_tokens.

Key bin_counts · version 0993c95c-e538-42eb-972d-9fbb21b27ab0

FOCUS

Fast Overlapping Token Combinations Using Sparsemax: an embedding initialisation for a replaced tokenizer's vocabulary.

Key focus_init · version 5a078e71-84ff-4039-9e32-7999a5f679f5

Bits per byte

Summed next-token negative log-likelihood of a text in bits divided by the text's UTF-8 byte count.

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

APT4 transplant of Qwen2.5-1.5B

Qwen2.5-1.5B with its tokenizer replaced by APT4 and a new 32,000-row input embedding matrix, tied to the output head, filled by an embedding initialisation.

Key qwen_apt4_transplant · version c8bc4afe-d6ab-4475-b668-4f01d4d149b2

Frequency-based Vocabulary Transfer

An embedding initialisation for a replaced tokenizer's vocabulary.

Key fvt_init · version 01087ca0-917c-46f8-a25e-ebc786018d72

Exact references

derived from

ff70da41-5976-4727-b9fa-0a20073709ff

Per-domain value and bin.

derived from

8dce818c-b288-4a4f-9cb8-0106815d8dd0

Per-domain value and bin.

derived from

013a25c0-5ae1-4dee-9f66-eeffe08773aa

Per-domain value and bin.

derived from

a86cf354-158d-4815-bee0-90e2132a7e3a

Per-domain value and bin.

derived from

4ad87e4a-3105-4067-a78e-badf5584dab8

Per-domain value and bin.

derived from

279fef10-7e48-4367-b4b0-17888a094abe

Per-domain value and bin.

derived from

e9922d93-0332-404f-82cd-5e536cf1f0b5

Per-domain value and bin.