Attributed research checkpoint
Seed: entity definitions, premise claims, and disputes where the seed paper's conclusions conflict with its own tables.
Assignment, status and accepted plans remain authoritative on each experiment.
Reported state
- Reported progress
- Checks of the paper that hold: "Table 1 arithmetic holds", "Context doubling holds". No experiment has run.
- Open issues
- Gaps: "No same-training control"; "APT4 recipe absent"; "No optimisation hyperparameters"; "FOCUS auxiliary space unspecified"; "Output-head initialisation unspecified"; "Recovery data unspecified"; "Token-count tokenizer unspecified"; "Post-training data unreleased"; "Multilingual subsets unlisted"; "Tokens-per-word protocol undefined". Thin evidence: "FOCUS choice borrowed from 1.5B"; "Fertility from one 232-word text"; "No variance information". Inconsistencies without a dispute: "Unmentioned regressions"; "Minitron rows imported unmarked"; "FVT citation unresolved". Entries of docs/seed-paper-notes.md in stw2/tokenizer-science-tax.
- Suggested next action
- Measure tokenizer fertility on scientific and engineering corpora (E1).
- Access needs
- Reading needs nothing. The gated Bielik v3 PL checkpoints, which carry the APT4 tokenizer files, need the SpeakLeash terms accepted once on Hugging Face.
Exact references
Exact record · P16
8368afe3-5846-42c2-9352-ffa0ff2db353Rests on one text: "Fertility from one 232-word text".
Exact record · P17
eed1981a-09e9-4aa9-bd57-3db32c7ba4a0Rests on one text: "Fertility from one 232-word text".
Exact record · P25
d664a095-2689-445e-ac40-a85c6202a460Composition unknown: "Recovery data unspecified".
Exact record · P26
4aeee3b3-fb89-4ae6-9e40-9668ffa62e84No arm adapts the original tokenizer on the same tokens: "No same-training control".
Exact record · P24
f5ada7c4-30c8-405f-8741-cd6fcfb58530Evidence reported elsewhere: "FOCUS choice borrowed from 1.5B".
Exact record · P30
0f7b83ab-0210-479b-88e4-ac7a40bbafa3The 7B model with the Polish tokenizer scores above its counterpart in English: evidence for "No same-training control".
Exact record · P29
36da68d9-97cf-4575-ad42-ed56e0e895e9Source of the dispute on English capabilities.
Exact record · P40
15132b6d-26e7-49b9-8217-b144a28236c7Disputed by the INCLUDE-base-44 and Belebele European averages.
Exact record · P35
cbcebd2f-77d5-48ae-9534-159d89911958Source of the dispute on preserved performance.
Exact record · P37
e0184bf2-dccb-469b-85be-cefff666bd50Named in the dispute on preserved performance.
Exact record · P34
e9fd6b0d-e665-41be-a6dc-290547a16735Regression not discussed by the paper: "Unmentioned regressions".
Exact record · P33
fdcd5f0a-2f62-43af-ae64-2289127e532aRegression not discussed by the paper: "Unmentioned regressions".