---
license: mit
pretty_name: SlayerLab Tokenizers
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
- config_name: benchmark
  data_files:
  - split: test
    path: results/tokenizer_benchmark.parquet
- config_name: evidence_review
  data_files:
  - split: review
    path: results/author_evidence_scores.csv
- config_name: leaderboard
  data_files:
  - split: test
    path: results/provisional_leaderboard.parquet
---

# SlayerLab Tokenizers

Normalized tokenizer artifacts collected from the contributor directories in
[`slayerlabs/tokenizer`](https://github.com/slayerlabs/tokenizer/tree/1a5cd2c2e4df2287b4c19b3dbf5051f5d460fdc1),
pinned to source commit `1a5cd2c2e4df2287b4c19b3dbf5051f5d460fdc1`.

The dataset contains one row per tokenizer: the 38 workshop submissions plus
the canonical SlayerLab Polish 32k tokenizer by `kacperwikiel`. Use the Dataset
Viewer to sort, filter, and compare tokenizers without navigating folders.

## Columns

- `author`: contributor's exact GitHub username, resolved from the source repository's pull-request history.
- `size`: tokenizer vocabulary size.
- `name`: original filename.
- `quick_status`: whether the artifact is ready to load or needs custom conversion.
- `hf_loadable`: whether it loads directly with `tokenizers.Tokenizer.from_file(...)`.
- `format`: normalized artifact-format classification.
- `model_type`: tokenizer model type reported by the source.
- `merge_count`: number of BPE merge rules when available.
- `normalizer`, `pre_tokenizer`, `decoder`: quick configuration summary.
- `unk_token`, `added_tokens_count`: special-token readiness indicators.
- `reported_metrics`: evaluation numbers reported by the author, encoded as JSON.
- `metadata`: normalized metadata encoded as a JSON string.
- `tokenizer_json`: complete, lossless original JSON file contents.
- `source_repo`: repository containing the original artifact.
- `source_path`: original path in the source repository.
- `source_commit`: pinned source Git commit.
- `bytes`: original JSON file size.
- `sha256`: checksum of the original JSON file.

## Formats

- `hf_tokenizers` (13 files): loadable with `tokenizers.Tokenizer.from_file(...)`.
- `custom_bpe` (17 files): custom BPE artifacts with `meta` and `model` fields.
- `custom_vocab` (1 file): vocabulary/merge mapping rather than a full tokenizer runtime file.
- `custom_experiment` (7 files): custom experiment or result JSON containing tokenizer data and metrics.

All source files were parsed as valid JSON. A file being valid JSON does not
imply it implements the Hugging Face Tokenizers serialization schema; check the
`format` column before loading `tokenizer_json`.

Reported metrics were produced with different texts and procedures, so they are
useful for inspecting an individual submission but not for ranking authors. A
fair quality ranking requires running every compatible tokenizer against the
same held-out Polish evaluation corpus.

## Common diagnostic benchmark

The `benchmark` configuration evaluates all 39 artifacts on the same versioned,
ten-domain Polish diagnostic suite. Fourteen run natively through the Rust
`tokenizers` library and 25 run through the included custom-format adapters.

- 39/39 artifacts executed successfully.
- 32 are reconstructed exactly; 7 are `core_only` because their intended
  pre-tokenizer was not serialized.
- 37/39 preserve exact input; two normalized `KateMajzel` variants convert a
  tab to a space.
- No unknown tokens were observed in this suite.

The benchmark reports compression, round-trip behavior, unknown-token rate,
per-domain results, serialized size, and local throughput. Throughput is only
comparable within the same `runtime`; Python reference adapters must not be
speed-ranked against native Rust tokenizers.

This is a small synthetic diagnostic corpus, not a statistically representative
held-out benchmark and not evidence of downstream language-model quality. Raw
compression must be compared within vocabulary-size bands.

## Evidence review

The `evidence_review` configuration separately scores artifact usability,
documentation, evaluation protocol, reproducibility, and claims discipline.
These scores judge the submitted evidence package—not tokenizer performance—and
must not be combined with compression metrics into a single winner score. See
`EVIDENCE_REVIEW.md` for the full evidence and limitations.

## Provisional scoring leaderboard

The `leaderboard` configuration provides a transparent 0–100 diagnostic quality
score. Eligibility requires successful execution, exact adapter fidelity, exact
round-trip on every suite record, and zero observed unknown tokens. Seven
`core_only` reconstructions and two round-trip failures remain visible but are
unranked.

The score adjusts compression for vocabulary size using author-balanced fits in
each domain, then combines 80% mean domain percentile with 20% lower-quartile
domain percentile. This rewards balanced performance and prevents an author
with many variants from defining the baseline. Scores within two points share
a rank tier.

Artifact readiness, evidence-package quality, traceability, file size, and speed
are displayed separately and do not influence the quality rank. `kacperwikiel`
is marked as a reference baseline. Best-of-many rows are explicitly labeled for
selection bias. See `LEADERBOARD_METHODOLOGY.md` for the complete formula and
limitations.