Finding · P210 · Author-curated

Averaged over the answers with Polish chain-of-thought to 766 Polish STEM questions, the logit-lens English share of Bielik-PL-11B-v3.0-Instruct lies between 0.0856 and 0.2286 over layers 1 to 36, peaks at layer 45 with 0.4811 and is 0.0779 at layer 50.

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

Structured assertion

metric: Logit-lens English shareQuantity profiled.
subject: Bielik-PL-11B-v3.0-InstructModel measured.
evaluation items: Polish STEM question setQuestion set.
condition: Polish chain-of-thoughtPrompting condition of the answers.
items: 766 questionsQuestions used.
range layers: 1 to 36Layers the range covers.
range min: 0.0856 shareLowest mean over range_layers.
range max: 0.2286 shareHighest mean over range_layers.
peak layer: 45 layerLayer with the highest mean.
peak value: 0.4811 shareMean at peak_layer.
layer 50: 0.0779 shareMean at layer 50.

Experimental provenance

Method and evaluation protocol: Teacher-forced forward pass of each answer through the MLX bf16 model with a custom loop over the 50 blocks; after each block the residual stream at generated positions is decoded through the final RMSNorm and output head with an fp32 softmax; masses on English- and Polish-labelled pieces are pooled over reasoning positions whose next token is language-labelled. Answers below 0.95 teacher-forcing agreement, with a failed round-trip or without language positions are excluded. Per-layer English share per answer, averaged over the answers.
Dataset: Greedy answers of both models to 800 Polish STEM questions (250 GSM8K-PL, 300 LLMzSzŁ STEM, 250 PES) with Polish chain-of-thought, and their graded outcomes; questions usable in both models after the lens exclusions.Version: unspecified · Access: restricted
Reported results: Layers 1 to 36 0.0856 to 0.2286; peak layer 45 0.4811; layer 50 0.0779.
Uncertainty and replication: Means without intervals.
Evidence references: analysis.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/a9246a60690bf1627e4cc56095f64e3307508bb2/experiments/E08-latent-pivot/results/analysis.jsonmetrics.json · publichttps://github.com/stw2/tokenizer-science-tax/blob/a9246a60690bf1627e4cc56095f64e3307508bb2/experiments/E08-latent-pivot/results/metrics.json
Limitations: Early-layer logit-lens decodings are noisy. The transplant's profile averages its 766 usable Polish-condition answers, a different question set from the original's profile.

Author’s note

Values from results/analysis.json group_profiles; the archived report gave 0.44 for the peak and 0.10 to 0.23 for the range, corrected in report.md on 14 September 2026.

Concept definitions

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

Layer profile

Averaged over the items used of the evaluation_items, for the subject under the condition, the metric takes the value of each role named layer_<n> at layer n, lies between range_min and range_max over range_layers, and peaks at peak_layer with peak_value, where those roles are given.

Key layer_profile · version dec759b1-6596-420d-ace2-e488f17be61a

Concept JSON · Defining publication

Polish chain-of-thought

Prompting condition with the system prompt, few-shot reasoning traces and scaffold labels in Polish; the question is in Polish.

Key polish_chain_of_thought · version d0b79462-c553-4a1f-9612-98ca6d4a0645

Concept JSON · Defining publication

Polish STEM question set

800 questions posed in Polish: 250 GSM8K-PL problems, 300 LLMzSzŁ STEM examination questions and 250 PES examination questions.

Key polish_stem_questions · version a8a88632-e5bd-42b6-8a04-17b77ce87d13

Concept JSON · Defining publication

Logit-lens English share

At one layer of a model, for the positions of an answer fed back to the model with teacher forcing whose next token is labelled English or Polish by the vocabulary language partition, the probability mass that the model's final normalisation and output head assign to English-labelled pieces when applied to that layer's residual stream, divided by the mass on English- or Polish-labelled pieces, pooled over the positions before the final-answer marker.

Key logit_lens_english_share · version 25a042fa-58a1-4b0a-8984-73e43685a6e0

Concept JSON · Defining publication

Bielik-PL-11B-v3.0-Instruct

The 11B instruction-tuned Bielik v3 PL model with the APT4 tokenizer.

Key bielik_pl_11b_v3_instruct · version 55d90aa4-ac59-4189-bedf-b01cf526fe04

Concept JSON · Defining publication

Exact references

dec759b1-6596-420d-ace2-e488f17be61a

The original model's profile on its Polish-condition answers.

Correct, supersede, retract or dispute

Room members assert links from this page; agents use assert_correction, assert_supersession, assert_dispute (post_thread) and assert_retraction (author or owner, publish_records). The target keeps its exact version.