tokenizers / LEADERBOARD_METHODOLOGY.md
kacperwikiel's picture
Add fair provisional tokenizer leaderboard
c8c3aff
|
Raw
History Blame Contribute Delete
3.98 kB

Provisional tokenizer leaderboard methodology

This leaderboard is a diagnostic ranking on the small synthetic suite, not a claim about downstream model quality.

Eligibility gates

A row receives a rank only when the benchmark succeeds, adapter fidelity is exact, every evaluation string round-trips exactly, and observed unknown-token rate is zero. core_only artifacts and round-trip failures remain in both outputs with a reason and no rank.

Vocabulary-adjusted provisional quality

Raw tokens per word favors large vocabularies. For each of the ten domains, the script fits a weighted least-squares line across eligible artifacts:

ln(domain tokens/word) = intercept + slope * log2(vocabulary size)

Each row is weighted by 1 / eligible rows from its author, so every author has equal total influence on the fitted baseline. This prevents one author's many variants from defining expected compression. Residuals are also converted to author-balanced percentiles, rescaled so the observed best is 100 and worst is 0. The adjusted_compression_index is the geometric mean residual ratio across domains: below 1 is better, but it is diagnostic rather than the score itself.

The quality score gives every domain equal influence:

0.80 * mean(domain percentile) + 0.20 * lower-quartile(domain percentiles)

The lower-quartile term rewards tokenizers that avoid weak domains. Quality is shown to one decimal. Adjacent entries within two points of the leading score in their tier share a competition rank, because this suite does not support fine-grained distinctions. This score is relative to the current eligible cohort and changes when submissions change.

Artifact readiness (reported separately)

Native Hugging Face Tokenizers artifacts receive 100. Exact custom artifacts requiring the Python reference adapter receive 70. This deliberately small column describes direct operational loadability. It is not included in the quality score.

Independent author evidence package (reported separately)

results/author_evidence_scores.csv is an independent author-level review of artifact usability, documentation, evaluation protocol, reproducibility, and claims discipline. Its total out of 20 is multiplied by five and exposed as evidence_package_score out of 100. Every tokenizer by an author inherits that author-level score. The kacperwikiel reference baseline was not part of the author evidence review and therefore has a null score. Evidence is not included in tokenizer quality or rank.

Source repository, path, pinned commit, and SHA-256 completeness remain visible as traceability_score and traceability_checks_json. These fields describe artifact provenance only and are not called evidence-package quality.

Encoding and decoding throughput are retained as info_only columns. They are not scored because native Rust and Python reference-adapter runtimes are not comparable. Unknown-token rate is also retained but is not a differentiator in the current suite, where every artifact reports zero observed unknowns.

kacperwikiel is explicitly labeled as the reference baseline. For authors with multiple eligible variants, their leading row is labeled author_best_of_N_selection_bias; choosing the best of many trials can inflate its apparent standing relative to single submissions.

Leave-one-author-out ranks are not reported. With only a tiny synthetic suite, few authors, uneven vocabulary-size coverage, and some authors occupying unique size regions, refitting after removing one author can become extrapolation and would look more authoritative than it is. Author-balanced fitting, explicit best-of-many labels, one-decimal scores, and two-point rank tiers are the current sensitivity safeguards. A larger natural held-out corpus and more authors per size band are needed before meaningful leave-one-author-out claims.

Reproduce after running the benchmark:

python3 build_leaderboard.py