Finding

Sign in with GitHub
← Publications

Finding · P173 · Author-curated

At time zero, the mean cross-entropy of APT4 transplants of Qwen2.5-1.5B on a 40-token Polish sentence is 8.8109 nats for FOCUS, 10.3366 for FVT and 13.7575 for random initialisation.

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

Structured assertion

Relation: Values by initialisation

metric
mean next-token cross-entropy in natsQuantity measured.
model
APT4 transplant of Qwen2.5-1.5BModels measured.
evaluation text
one Polish sentence of 40 APT4 tokens about tokenization and scientific abilityText measured.
setting
time zero: no training after embedding initialisationTraining state of the models.
focus value
8.8109 natsValue for FOCUS initialisation.
fvt value
10.3366 natsValue for FVT initialisation.
random value
13.7575 natsValue for random initialisation.

Experimental provenance

Method and evaluation protocol
Forward pass of each freshly built transplant on one sentence during the build, with the APT4 tokenizer loaded from its own directory.
Dataset
One Polish sentence fixed in scripts/02_build_inits.py.Version: unspecified · Access: public
Reported results
FOCUS 8.8109, FVT 10.3366, random 13.7575.
Uncertainty and replication
Single sentence; no interval.
Limitations
One sentence; a sanity check of the builds, not an endpoint. The build statistics were copied from the restricted model directories into the repository in the port.

Author’s note

Values from results/build_stats/<init>.json, sanity_pl_loss.

Concept definitions

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

Values by initialisation

The metric of the model on the evaluation text in the setting is focus_value, fvt_value and random_value for FOCUS, FVT and random initialisation.

Key values_by_initialisation · version 4f7354da-9088-4141-85b3-f15a5312e066

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

Exact references

related

2174054b-06f0-40a7-b2f7-96cecb0b7c75

The same ordering on one sentence, before per-domain scoring.