On 598 translation-paired Belebele items, the English likelihood multiple-choice accuracy of Qwen2.5-1.5B is 0.6087 after 500M tokens of Polish-heavy continued pretraining with its original tokenizer and 0.7458 before it, a difference of -0.1371 (95% interval -0.1739 to -0.1020).

Relation: Paired difference

metric

Likelihood multiple-choice accuracyQuantity compared.

language

EnglishLanguage of the items the accuracy is measured on.

benchmark

Translation-paired Belebele itemsItem pairs scored.

subject

Qwen2.5-1.5B continued with its own tokenizerModel after continued pretraining with its original tokenizer.

subject value

0.6087 accuracyMetric of the subject.

comparator

Qwen2.5-1.5BModel before continued pretraining.

comparator value

0.7458 accuracyMetric of the comparator.

value

-0.1371 accuracySubject minus comparator.

interval low

-0.1739 accuracyLower bound of the paired bootstrap 95% interval.

interval high

-0.1020 accuracyUpper bound of the paired bootstrap 95% interval.

training

500,170,752 Qwen tokens of continued pretraining of each arm on the same document sequence, constant learning rate before any decay anneal, one run per armTraining the continued-pretraining arms received.

Method and evaluation protocol

Item-level paired bootstrap over the analysed pairs (2000 resamples, numpy PCG64 seed 20260703): every resample draws the same pairs for both models and both languages; the estimate is the difference on all pairs.

Dataset

598 pairs of the same FLORES passage, question and four options in English and in Polish from the Belebele test split: a seed-42 sample of 600 with the structurally malformed pairs excluded.Version: facebook/belebele@7899cdfa4e1e0d733fd77c848e2c273cb1d32be2 · Access: public

Reported results

0.6087 against 0.7458: -0.1371 [-0.1739, -0.1020].

Uncertainty and replication

Paired bootstrap percentile 95% interval; one training run per checkpoint, so training variance is not included.

Evidence references

analysis.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/5e8117ffb3f168331ee63f1032cd40a62c6296bd/experiments/E11-cross-language-access/results/analysis.jsonmetrics.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/5e8117ffb3f168331ee63f1032cd40a62c6296bd/experiments/E11-cross-language-access/results/metrics.json

Limitations

One model family at 1.5B parameters; the checkpoints are pre-anneal, from one training run each. English MMLU items may be in Qwen2.5 pretraining data, so MMLU gaps can be inflated; Polish MMLU items are machine translations. The 13 excluded pairs were chosen after the base model's cells had been scored. The two models compared may favour different option letters, so the difference can reflect letter priors as well as item knowledge. Letter priors: on Polish MMLU items the checkpoint that kept the original tokenizer predicts A for 4 of 600 items and the APT4 checkpoint predicts B for 437, against 134 to 160 gold items per letter; on Polish Belebele items the APT4 checkpoint predicts B for 360 and the other checkpoint D for 326 of 600; the base model's predictions are the least concentrated (all scored rows, results/metrics.json pred_count_*).