Two runs of the paired likelihood scoring of Qwen2.5-1.5B on the first 10 items of each of the four item files produce byte-identical score files of 40 rows.

Relation: Byte-identical repeat

procedure

likelihood multiple-choice scoring of the first 10 items of each item file, torch MPS bf16Procedure repeated.

subject

Qwen2.5-1.5BModel scored.

runs

2 runsNumber of runs compared.

rows per run

40 rowsRows in each score file.

Method and evaluation protocol

Run the scoring script twice with the same arguments and compare the sha256 of the two score files.

Dataset

First 10 items of each of the four translation-paired item files (Belebele and MMLU, English and Polish).Version: unspecified · Access: public

Reported results

Both files sha256 7f3234fd173d904874e4e36e03d27cebdcd2dc74f636b6f93aafae096b79689d.

Uncertainty and replication

Exact byte comparison.

Evidence references

smoke_run1.jsonl · publichttps://github.com/stw2/tokenizer-science-tax/blob/5e8117ffb3f168331ee63f1032cd40a62c6296bd/experiments/E11-cross-language-access/results/smoke_run1.jsonlsmoke_run2.jsonl · publichttps://github.com/stw2/tokenizer-science-tax/blob/5e8117ffb3f168331ee63f1032cd40a62c6296bd/experiments/E11-cross-language-access/results/smoke_run2.jsonl

Limitations

The design's smoke gate names all three models; it ran on the base model only. Byte identity on one machine says nothing about other hardware.