Add fair provisional tokenizer leaderboard

LEADERBOARD_METHODOLOGY.md

@@ -0,0 +1,84 @@
+# 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:
+
+```bash
+python3 build_leaderboard.py
+```

README.md

@@ -14,6 +14,10 @@ configs:
   data_files:
   - split: review
     path: results/author_evidence_scores.csv
+- config_name: leaderboard
+  data_files:
+  - split: test
+    path: results/provisional_leaderboard.parquet
 ---
 
 # SlayerLab Tokenizers
@@ -92,3 +96,23 @@ 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.

build_leaderboard.py

@@ -0,0 +1,218 @@
+#!/usr/bin/env python3
+"""Build the transparent provisional leaderboard from benchmark results."""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import math
+import statistics
+from collections import Counter
+from pathlib import Path
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+
+MEAN_DOMAIN_WEIGHT = 0.80
+LOWER_QUARTILE_WEIGHT = 0.20
+TIE_WINDOW_POINTS = 2.0
+
+
+def args_parse() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--results", type=Path, default=Path("results/tokenizer_benchmark.parquet"))
+    parser.add_argument("--dataset", type=Path, default=Path("data/train-00000-of-00001.parquet"))
+    parser.add_argument(
+        "--author-evidence", type=Path, default=Path("results/author_evidence_scores.csv")
+    )
+    parser.add_argument("--csv", type=Path, default=Path("results/provisional_leaderboard.csv"))
+    parser.add_argument("--parquet", type=Path, default=Path("results/provisional_leaderboard.parquet"))
+    return parser.parse_args()
+
+
+def weighted_ols_residuals(points: list[tuple[float, float]], weights: list[float]) -> list[float]:
+    """Residuals for weighted y = intercept + slope*x."""
+    total_weight = sum(weights)
+    x_mean = sum(weight * point[0] for point, weight in zip(points, weights)) / total_weight
+    y_mean = sum(weight * point[1] for point, weight in zip(points, weights)) / total_weight
+    denominator = sum(weight * (x - x_mean) ** 2 for (x, _), weight in zip(points, weights))
+    slope = sum(
+        weight * (x - x_mean) * (y - y_mean)
+        for (x, y), weight in zip(points, weights)
+    ) / denominator
+    intercept = y_mean - slope * x_mean
+    return [y - (intercept + slope * x) for x, y in points]
+
+
+def author_weighted_percentile_scores(values: list[float], weights: list[float]) -> list[float]:
+    """Author-balanced percentile, rescaled so observed best=100 and worst=0."""
+    if len(values) == 1:
+        return [100.0]
+    raw = []
+    total = sum(weights)
+    for value in values:
+        better = sum(weight for candidate, weight in zip(values, weights) if candidate < value)
+        tied = sum(weight for candidate, weight in zip(values, weights) if candidate == value)
+        raw.append(100.0 * (1.0 - (better + tied / 2) / total))
+    low, high = min(raw), max(raw)
+    return [100.0 * (score - low) / (high - low) for score in raw]
+
+
+def eligibility(row: dict) -> tuple[bool, str]:
+    if row["status"] != "ok":
+        return False, "benchmark_error"
+    if row["adapter_fidelity"] != "exact":
+        return False, "core_only_adapter"
+    if not row["roundtrip_pass"]:
+        return False, "roundtrip_failure"
+    if row["unk_rate"] != 0:
+        return False, "nonzero_unk_rate"
+    return True, "eligible"
+
+
+def traceability_score(source: dict) -> tuple[float, str]:
+    checks = {
+        "source_repo": bool(source.get("source_repo")),
+        "source_path": bool(source.get("source_path")),
+        "source_commit": bool(source.get("source_commit")),
+        "sha256": len(source.get("sha256") or "") == 64,
+    }
+    return 25.0 * sum(checks.values()), json.dumps(checks, sort_keys=True)
+
+
+def main() -> None:
+    args = args_parse()
+    benchmark = pq.read_table(args.results).to_pylist()
+    sources = {row["sha256"]: row for row in pq.read_table(args.dataset).to_pylist()}
+    with args.author_evidence.open(newline="", encoding="utf-8") as handle:
+        author_evidence = {row["author"]: row for row in csv.DictReader(handle)}
+    domains = sorted(json.loads(benchmark[0]["domain_metrics_json"]))
+    eligible_indices = [i for i, row in enumerate(benchmark) if eligibility(row)[0]]
+    eligible_author_counts = Counter(benchmark[i]["author"] for i in eligible_indices)
+    author_weights = [1.0 / eligible_author_counts[benchmark[i]["author"]] for i in eligible_indices]
+
+    # Fit the expected log(tokens/word) vs log2(vocabulary size) relation in
+    # each domain. Averaging residuals gives every domain equal influence.
+    residuals_by_index = {index: [] for index in eligible_indices}
+    domain_scores_by_index = {index: [] for index in eligible_indices}
+    for domain in domains:
+        points = []
+        for index in eligible_indices:
+            row = benchmark[index]
+            tpw = json.loads(row["domain_metrics_json"])[domain]["tokens_per_word"]
+            points.append((math.log2(row["size"]), math.log(tpw)))
+        residuals = weighted_ols_residuals(points, author_weights)
+        percentiles = author_weighted_percentile_scores(residuals, author_weights)
+        for index, residual, percentile in zip(eligible_indices, residuals, percentiles):
+            residuals_by_index[index].append(residual)
+            domain_scores_by_index[index].append(percentile)
+
+    adjusted = {
+        index: math.exp(sum(values) / len(values))
+        for index, values in residuals_by_index.items()
+    }
+    quality_scores = {}
+    mean_domain_scores = {}
+    lower_quartile_scores = {}
+    for index, scores in domain_scores_by_index.items():
+        mean_domain_scores[index] = statistics.mean(scores)
+        lower_quartile_scores[index] = statistics.quantiles(scores, n=4, method="inclusive")[0]
+        quality_scores[index] = (
+            MEAN_DOMAIN_WEIGHT * mean_domain_scores[index]
+            + LOWER_QUARTILE_WEIGHT * lower_quartile_scores[index]
+        )
+
+    rows = []
+    for index, source_result in enumerate(benchmark):
+        is_eligible, reason = eligibility(source_result)
+        source = sources[source_result["sha256"]]
+        traceability, traceability_detail = traceability_score(source)
+        reviewed_evidence = author_evidence.get(source_result["author"])
+        evidence_package_score = (
+            float(reviewed_evidence["total"]) * 5 if reviewed_evidence is not None else None
+        )
+        readiness = 100.0 if source_result["adapter_status"] == "native" else 70.0
+        quality = quality_scores.get(index)
+        rows.append({
+            "rank": None,
+            "eligible": is_eligible,
+            "eligibility_reason": reason,
+            "author": source_result["author"],
+            "name": source_result["name"],
+            "source_path": source_result["source_path"],
+            "vocab_size": source_result["size"],
+            "tokens_per_word": source_result["tokens_per_word"],
+            "adjusted_compression_index": adjusted.get(index),
+            "mean_domain_percentile": mean_domain_scores.get(index),
+            "lower_quartile_domain_percentile": lower_quartile_scores.get(index),
+            "provisional_quality_score": round(quality, 1) if quality is not None else None,
+            "artifact_readiness_score": readiness,
+            "evidence_package_score": evidence_package_score,
+            "evidence_package_total_20": (
+                int(reviewed_evidence["total"]) if reviewed_evidence is not None else None
+            ),
+            "evidence_package_judgment": (
+                reviewed_evidence["evidence_judgment"] if reviewed_evidence is not None else ""
+            ),
+            "traceability_score": traceability,
+            "reference_baseline": source_result["author"] == "kacperwikiel",
+            "author_eligible_submission_count": eligible_author_counts.get(source_result["author"], 0),
+            "selection_bias_label": "",
+            "adapter_status": source_result["adapter_status"],
+            "adapter_fidelity": source_result["adapter_fidelity"],
+            "runtime": source_result["runtime"],
+            "roundtrip_pass": source_result["roundtrip_pass"],
+            "unk_rate": source_result["unk_rate"],
+            "encode_mb_per_s_info_only": source_result["encode_mb_per_s"],
+            "decode_mb_per_s_info_only": source_result["decode_mb_per_s"],
+            "traceability_checks_json": traceability_detail,
+            "sha256": source_result["sha256"],
+        })
+
+    ranked = sorted(
+        (row for row in rows if row["eligible"]),
+        key=lambda row: (-row["provisional_quality_score"], row["source_path"]),
+    )
+    author_best_path = {}
+    for row in ranked:
+        author_best_path.setdefault(row["author"], row["source_path"])
+    tier_start = 1
+    tier_anchor = None
+    for position, row in enumerate(ranked, 1):
+        score = row["provisional_quality_score"]
+        if tier_anchor is None or tier_anchor - score > TIE_WINDOW_POINTS:
+            tier_start, tier_anchor = position, score
+        row["rank"] = tier_start
+        count = row["author_eligible_submission_count"]
+        if count == 1:
+            row["selection_bias_label"] = "single_submission"
+        elif row["source_path"] == author_best_path[row["author"]]:
+            row["selection_bias_label"] = f"author_best_of_{count}_selection_bias"
+        else:
+            row["selection_bias_label"] = f"variant_among_{count}"
+    rows.sort(key=lambda row: (
+        not row["eligible"],
+        row["rank"] or 10**9,
+        -(row["provisional_quality_score"] or -1),
+        row["source_path"],
+    ))
+
+    args.csv.parent.mkdir(parents=True, exist_ok=True)
+    with args.csv.open("w", newline="", encoding="utf-8") as handle:
+        writer = csv.DictWriter(handle, fieldnames=list(rows[0]))
+        writer.writeheader()
+        writer.writerows(rows)
+    pq.write_table(pa.Table.from_pylist(rows), args.parquet, compression="zstd")
+    print(f"eligible={len(ranked)} unranked={len(rows)-len(ranked)}")
+    for row in ranked[:10]:
+        print(
+            f"{row['rank']:2}. {row['author']:16} size={row['vocab_size']:6} "
+            f"quality={row['provisional_quality_score']:.1f} adjusted={row['adjusted_compression_index']:.4f} "
+            f"{row['source_path']}"
+        )
+
+
+if __name__ == "__main__":
+    main()

results/provisional_leaderboard.csv

@@ -0,0 +1,40 @@
+rank,eligible,eligibility_reason,author,name,source_path,vocab_size,tokens_per_word,adjusted_compression_index,mean_domain_percentile,lower_quartile_domain_percentile,provisional_quality_score,artifact_readiness_score,evidence_package_score,evidence_package_total_20,evidence_package_judgment,traceability_score,reference_baseline,author_eligible_submission_count,selection_bias_label,adapter_status,adapter_fidelity,runtime,roundtrip_pass,unk_rate,encode_mb_per_s_info_only,decode_mb_per_s_info_only,traceability_checks_json,sha256
+1,True,eligible,Maggio333,lektury-naive.json,Arek/1024/lektury-naive.json,1024,3.792079207920792,0.9389787490912972,71.91451906597675,54.95269618353344,68.5,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,author_best_of_11_selection_bias,custom_adapter,exact,python_reference_adapter,True,0.0,0.3045090581016393,50.125886930647745,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",43d95eadb72c93f7fa8daadc09007d414ea378cf0c2a79c1543b1cf0fdfbea47
+2,True,eligible,dawidmajewski,wikinews_5000.tokenizer.json,dawidm/vocabs/wikinews_5000.tokenizer.json,5000,2.9257425742574257,0.9587331767938928,69.68469678483488,43.78537735849056,64.5,100.0,65.0,13,Seven loadable artifacts and concrete corpus tables; primarily exploratory evidence.,100.0,False,7,author_best_of_7_selection_bias,native,exact,rust_tokenizers,True,0.0,16.344851851456966,4.966517314776119,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",c137d9111598ea4889277f3a52f73e164252d6455c15deba38fbe31982256b08
+2,True,eligible,janbanot,tokenizer.json,Janek/tokenizer.json,8192,2.732673267326733,0.9732298433771442,65.90512051177087,50.188679245283005,62.8,100.0,80.0,16,Broadest qualitative and stress analysis with a directly loadable final artifact.,100.0,False,1,single_submission,native,exact,rust_tokenizers,True,0.0,5.490510692706898,5.782178217821782,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",85b81524047f486ed36cee3174622b3d473bc1934b13b45999b3f30d5d8042f2
+4,True,eligible,Maggio333,slayer-v1.json,Arek/32000/slayer-v1.json,32000,2.123762376237624,0.9584209387960462,63.0812734997565,37.35253407006222,57.9,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,variant_among_11,custom_adapter,exact,python_reference_adapter,True,0.0,1.0930808623145933,68.13333333333333,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",f88ed0269ea30781d2d15effbc22aeb29eae7476233d307e1f62bbcc42fb8f56
+4,True,eligible,dawidmajewski,dynaword-100-novels-pretok_1500.tokenizer.json,dawidm/vocabs/dynaword-100-novels-pretok_1500.tokenizer.json,1500,3.608910891089109,0.9528767264978109,61.84506034846827,37.116084190966255,56.9,100.0,65.0,13,Seven loadable artifacts and concrete corpus tables; primarily exploratory evidence.,100.0,False,7,variant_among_7,native,exact,rust_tokenizers,True,0.0,4.973236009732361,4.422646952758745,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",a5aa6afc010d87edbf35322cd52884ec2eff3abbbe584463b902dd2f02e10069
+4,True,eligible,Maggio333,lektury-fast.json,Arek/1024/lektury-fast.json,1024,3.876237623762376,0.9586803136190613,60.568536041587826,37.89455482897383,56.0,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,variant_among_11,custom_adapter,exact,python_reference_adapter,True,0.0,1.4142614381738037,47.84345546470258,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",9831627c449df6ec3cc898a134370d191128a40f0709fd2a304b9efc4db3166d
+7,True,eligible,Maggio333,lektury-naive.json,Arek/512/lektury-naive.json,512,4.4504950495049505,0.9808549779338848,57.916922447790625,40.07004371724766,54.3,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,variant_among_11,custom_adapter,exact,python_reference_adapter,True,0.0,0.36971313829779545,43.28552066862435,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",ca2be8a6ff6d62a1924efb1d46040f412a0f03d23cf744e76e1dc15599754f61
+7,True,eligible,olajachymiak,quo_vadis_tokenizer.json,ola/quo_vadis_tokenizer.json,512,4.46039603960396,0.9805826167609321,57.35455646012651,40.525237123498556,54.0,70.0,75.0,15,Clear controlled experimental progression and candid overfitting analysis.,100.0,False,4,author_best_of_4_selection_bias,custom_adapter,exact,python_reference_adapter,True,0.0,0.37403928266438946,43.59324347380993,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",7ee222c61a6ae50871f334042453b3a28bb3e9ab40abb54a552c1b50064f4441
+9,True,eligible,olajachymiak,polish_bpe_8k.json,ola/polish_bpe_8k.json,8000,2.8217821782178216,0.984768071148563,55.547504460367044,34.90056818181817,51.4,70.0,75.0,15,Clear controlled experimental progression and candid overfitting analysis.,100.0,False,4,variant_among_4,custom_adapter,exact,python_reference_adapter,True,0.0,1.2140564453092517,59.82439024390244,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",e7478670752572ecdb17ae6b83ad11918a9d4f84d043273df01ad20e6836c555
+9,True,eligible,Maggio333,lektury-fast.json,Arek/512/lektury-fast.json,512,4.475247524752476,0.9831085899896038,53.43763741529055,37.212826972555725,50.2,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,variant_among_11,custom_adapter,exact,python_reference_adapter,True,0.0,1.633675236445984,41.33969743548257,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",1548ca5d178e248b3f6896e9cb37a0d4f012ebef7753bf01e625bd6f6397c00e
+9,True,eligible,olajachymiak,diverse_tokenizer.json,ola/diverse_tokenizer.json,512,4.564356435643564,1.0057429707600447,53.07205990808532,38.865566037735846,50.2,70.0,75.0,15,Clear controlled experimental progression and candid overfitting analysis.,100.0,False,4,variant_among_4,custom_adapter,exact,python_reference_adapter,True,0.0,0.3848415120731787,42.04723113634494,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",cbd552cffd0dc16a3c2ed22c50f2e42bb361f193b09366f0bc51d8e0d056e0f5
+12,True,eligible,kacperwikiel,polish_bpe_32k.json,tokenizers/polish_bpe_32k.json,32768,2.1881188118811883,0.9881875363429086,54.193107494714084,29.634877287558517,49.3,100.0,,,,100.0,True,1,single_submission,native,exact,rust_tokenizers,True,0.0,6.712643678160919,6.7471215802260485,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",3f1fc177f82e2ba5ec9881c80ca753535df610134bde3722e5653c362c70dd39
+12,True,eligible,KateMajzel,tokenizer z boilerplate.json,KasiaMP/wyniki/tokenizer z boilerplate.json,2256,3.4554455445544554,0.9829756709565203,50.995092957262464,37.130267399853665,48.2,70.0,90.0,18,Strongest controlled-methodology package; excellent failure analysis and explicit limits.,100.0,False,5,author_best_of_5_selection_bias,custom_adapter,exact,python_reference_adapter,True,0.0,0.2851473921848038,51.82030220058817,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",0da72d2b6e05c5f66dbf985fe8db097db0641e8a9c85e919f3f30b86fef5e196
+14,True,eligible,Maggio333,lektury-naive.json,Arek/2048/lektury-naive.json,2048,3.48019801980198,0.9704790416884974,50.368783678592564,24.630681818181806,45.2,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,variant_among_11,custom_adapter,exact,python_reference_adapter,True,0.0,0.275860033693433,52.862068965517246,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",c688ef26779b85b8ac4be7fee644a17e8c795725dd34879693af60cacd4a62ed
+14,True,eligible,Maggio333,lektury-fast.json,Arek/2048/lektury-fast.json,2048,3.51980198019802,0.9792948698311722,51.222674443807755,20.543558520936934,45.1,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,variant_among_11,custom_adapter,exact,python_reference_adapter,True,0.0,1.32369131138694,52.18723404255319,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",7c6a5b0e1cbebf257ce21381f87eff88a03da61e867795777522c076671bd0a0
+14,True,eligible,dawidmajewski,wikinews-pretok_16000.tokenizer.json,dawidm/vocabs/wikinews-pretok_16000.tokenizer.json,16000,2.504950495049505,0.9978697749021642,52.586614562228526,14.254376327094944,44.9,100.0,65.0,13,Seven loadable artifacts and concrete corpus tables; primarily exploratory evidence.,100.0,False,7,variant_among_7,native,exact,rust_tokenizers,True,0.0,5.268801682717093,7.183154027814223,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",7ba85edcc1de57bd5636957d5abea6c39831ffc32a4cebb17f3df81584c265b7
+17,True,eligible,ktalik,tokenizer.json,Konrad/tokenizer.json,456,4.707920792079208,1.017081882000958,44.23608493235591,24.952830188679233,40.4,100.0,50.0,10,Loadable minimal tokenizer and plots but sparse protocol and missing training code.,100.0,False,1,single_submission,native,exact,rust_tokenizers,True,0.0,13.427698021319646,3.717868809918197,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",2aa1dbeec838c8935e3f37f209a772d5f9cd2f776a784c26f1676332877b880f
+17,True,eligible,KateMajzel,hf_bez_regex.json,KasiaMP/wyniki/hf_bez_regex.json,6756,2.9554455445544554,1.0138025121771361,43.984780848129304,20.00462528961649,39.2,100.0,90.0,18,Strongest controlled-methodology package; excellent failure analysis and explicit limits.,100.0,False,5,variant_among_5,native,exact,rust_tokenizers,True,0.0,13.205954308948694,5.542619963555375,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",fe94408fa70d169a0b6da197ec1d4a3798fa16a15cedc2b9c447c892dac9182e
+17,True,eligible,dawidmajewski,wikinews_16000.tokenizer.json,dawidm/vocabs/wikinews_16000.tokenizer.json,16000,2.49009900990099,1.0040150144609825,44.74422585583652,16.634433962264133,39.1,100.0,65.0,13,Seven loadable artifacts and concrete corpus tables; primarily exploratory evidence.,100.0,False,7,variant_among_7,native,exact,rust_tokenizers,True,0.0,14.859546749898223,5.825220583206037,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",87e9b00dea6090df47148e2c77e97d6426f18875f871d088c32164e647f8af90
+17,True,eligible,KateMajzel,wyniki FINALNY.json,KasiaMP/wyniki/wyniki FINALNY.json,6756,2.9603960396039604,1.0162484739107795,43.13949782926138,20.00462528961649,38.5,70.0,90.0,18,Strongest controlled-methodology package; excellent failure analysis and explicit limits.,100.0,False,5,variant_among_5,custom_adapter,exact,python_reference_adapter,True,0.0,0.24253612095353874,54.586241276171485,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",cdd61ab6f7ee1722197fe33708390ed31291c35ae85cb1b7cb26edd13fdea87a
+17,True,eligible,KateMajzel,wyniki.json,KasiaMP/wyniki/wyniki.json,6756,2.9603960396039604,1.0162484739107795,43.13949782926138,20.00462528961649,38.5,70.0,90.0,18,Strongest controlled-methodology package; excellent failure analysis and explicit limits.,100.0,False,5,variant_among_5,custom_adapter,exact,python_reference_adapter,True,0.0,0.24286591282650455,63.21649484536083,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",e46e7772bdc2912c5aa717ca891036e87ed881c6d65e1117dcefbc6f2d3b6ba2
+22,True,eligible,KateMajzel,hf_z_regex.json,KasiaMP/wyniki/hf_z_regex.json,6756,2.985148514851485,1.0134746776130548,43.20817071034978,15.860087007290954,37.7,100.0,90.0,18,Strongest controlled-methodology package; excellent failure analysis and explicit limits.,100.0,False,5,variant_among_5,native,exact,rust_tokenizers,True,0.0,6.6184565569347,6.6184565569347,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",f1203515379795dca3e8f4228154b65edfc8c0eb00892227a77e0ffc2144537e
+22,True,eligible,Maggio333,slayer-v1.json,Arek/64000/slayer-v1.json,64000,1.995049504950495,1.0182208604702319,41.13184246567055,14.408170740423085,35.8,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,variant_among_11,custom_adapter,exact,python_reference_adapter,True,0.0,1.110634561382664,69.94570424784415,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",f69705ed9b8f38315330dd01efd37a608fbad828f360e1731c88b925b241b6dc
+24,True,eligible,Maggio333,lektury-fast.json,Arek/4096/lektury-fast.json,4096,3.2475247524752477,1.0175788877547602,39.296628349600496,11.384735972654095,33.7,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,variant_among_11,custom_adapter,exact,python_reference_adapter,True,0.0,1.2923981193204668,53.63140218303947,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",560bdf65c10a6029fb628767423808f42cae91f873f9a3c2c2f4b183d4df7a08
+25,True,eligible,dawidmajewski,syzyfowe-prace_5000.tokenizer.json,dawidm/vocabs/syzyfowe-prace_5000.tokenizer.json,5000,3.1683168316831685,1.0375276214720952,33.22404212097797,13.726325757575744,29.3,100.0,65.0,13,Seven loadable artifacts and concrete corpus tables; primarily exploratory evidence.,100.0,False,7,variant_among_7,native,exact,rust_tokenizers,True,0.0,13.941560036013424,5.343790849673202,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",5118904ba059b718853da1853e6bca9cbcda7cbf50dd88ae2297a742b011668f
+25,True,eligible,Maggio333,lektury-fast.json,Arek/8192/lektury-fast.json,8192,2.9752475247524752,1.0456938536221563,33.883705138235655,5.0225128644939705,28.1,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,variant_among_11,custom_adapter,exact,python_reference_adapter,True,0.0,1.2302541401703257,56.17236451577443,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",3222cbb0a14c2af56163da27c798509d7c68056c1b38d994a493ef83c8f41456
+25,True,eligible,olajachymiak,quo_vadis_tokenizer_4k.json,ola/quo_vadis_tokenizer_4k.json,4096,3.282178217821782,1.0355522606817549,31.588797465012583,10.88332514524949,27.4,70.0,75.0,15,Clear controlled experimental progression and candid overfitting analysis.,100.0,False,4,variant_among_4,custom_adapter,exact,python_reference_adapter,True,0.0,0.2637891792749463,51.60052509340604,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",bcaee7aa2023db58e6eba6970f703535a2d2cccde1f6f8d4f2d25b92e6520c83
+28,True,eligible,Maggio333,lektury-fast.json,Arek/15000/lektury-fast.json,15000,2.782178217821782,1.0801890598917583,32.90631740679627,1.6997212692967256,26.7,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,variant_among_11,custom_adapter,exact,python_reference_adapter,True,0.0,1.2041634363195488,58.493589743589745,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",98d24e6c275c2629da620fa11b28ecca4bafe46c1586f68e7577f47cae5a028e
+28,True,eligible,dawidmajewski,syzyfowe-prace-pretok_5000.tokenizer.json,dawidm/vocabs/syzyfowe-prace-pretok_5000.tokenizer.json,5000,3.222772277227723,1.0444674592023862,30.094730407147697,10.121486784516296,26.1,100.0,65.0,13,Seven loadable artifacts and concrete corpus tables; primarily exploratory evidence.,100.0,False,7,variant_among_7,native,exact,rust_tokenizers,True,0.0,5.924637681159421,6.054311080218635,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",b1f5fae33d469136c86f64ac0fa2f5e1d552a35a8d735cc3f083b873c173bc65
+30,True,eligible,dawidmajewski,syzyfowe-prace_16000.tokenizer.json,dawidm/vocabs/syzyfowe-prace_16000.tokenizer.json,16000,2.8613861386138613,1.1538675877907254,13.922171132902932,0.0,11.1,100.0,65.0,13,Seven loadable artifacts and concrete corpus tables; primarily exploratory evidence.,100.0,False,7,variant_among_7,native,exact,rust_tokenizers,True,0.0,16.038249079343824,5.53929539295393,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",9d42cf63df01609812a8c5530d0e3dfd8577e120357b82188e8b4d74316d3a56
+,False,core_only_adapter,Maggio333,slayer-v2.json,Arek/115200/slayer-v2.json,115200,1.9405940594059405,,,,,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,,custom_adapter,core_only,python_reference_adapter,True,0.0,0.20842136518849286,80.33328093067128,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",45c33b69a948a544414bd9271b99612c06907e0b3058303775355660370cd740
+,False,core_only_adapter,Maggio333,slayer-v2.json,Arek/128000/slayer-v2.json,128000,1.9108910891089108,,,,,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,,custom_adapter,core_only,python_reference_adapter,True,0.0,0.21379509727258542,74.17981225200813,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",76b23bed57f44910ca7a20afb9853703fdcdc3bca517b446bb9b0b2b8473a81b
+,False,core_only_adapter,Maggio333,slayer-v2.json,Arek/256000/slayer-v2.json,256000,1.7920792079207921,,,,,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,,custom_adapter,core_only,python_reference_adapter,True,0.0,0.20341344479275517,76.17391304347827,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",f4995613851f46f62eac4887056bd30aec2f60617e6fad7950f9aaec0127c73c
+,False,core_only_adapter,Maggio333,slayer-v2.json,Arek/32000/slayer-v2.json,32000,2.1534653465346536,,,,,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,,custom_adapter,core_only,python_reference_adapter,True,0.0,0.2229737100469074,67.01639344262296,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",b845c906a06d1c267d19d4b06010f0ea687794221ae77f60a2cdd2353da054b2
+,False,core_only_adapter,Maggio333,slayer-v2.json,Arek/512000/slayer-v2.json,512000,1.683168316831683,,,,,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,,custom_adapter,core_only,python_reference_adapter,True,0.0,0.19854727633416194,82.12353350833021,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",6e1451eeda2d8840081d4e42307c34ebf108109d7f3b848d2d855192b3dc175b
+,False,core_only_adapter,Maggio333,slayer-v2.json,Arek/64000/slayer-v2.json,64000,2.014851485148515,,,,,70.0,75.0,15,Deepest research narrative and strong caveats; missing runtime and scripts limit replay.,100.0,False,11,,custom_adapter,core_only,python_reference_adapter,True,0.0,0.2184590330173824,71.30232558139535,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",25e874b897f5011f00863290613e1d80ae94d410b90d94070efeaa32c60bb987
+,False,roundtrip_failure,KateMajzel,hf_norm_filtrowany.json,KasiaMP/wyniki/hf_norm_filtrowany.json,6756,2.98019801980198,,,,,100.0,90.0,18,Strongest controlled-methodology package; excellent failure analysis and explicit limits.,100.0,False,5,,native,exact,rust_tokenizers,False,0.0,6.851396648044694,5.467677218011591,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",bccc567eb0d55f9f5d2bbbcdb9e8af2c080c005b5f92490b04713a24efba4928
+,False,roundtrip_failure,KateMajzel,hf_normalizowany.json,KasiaMP/wyniki/hf_normalizowany.json,6756,2.9306930693069306,,,,,100.0,90.0,18,Strongest controlled-methodology package; excellent failure analysis and explicit limits.,100.0,False,5,,native,exact,rust_tokenizers,False,0.0,5.7740112994350286,5.871673484395826,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",152f8768c2ccc3f61c80c2ac421a2fa58835f79df56506bf6f4d4378cfba944e
+,False,core_only_adapter,p4pryk,polish_bpe_vocab.json,patryk/polish_bpe_vocab.json,16000,2.4158415841584158,,,,,70.0,45.0,9,Useful design explanation but only a custom vocabulary and merge map was submitted.,100.0,False,0,,custom_adapter,core_only,python_reference_adapter,True,0.0,0.22690101757631823,58.67942583732057,"{""sha256"": true, ""source_commit"": true, ""source_path"": true, ""source_repo"": true}",90302fed4b3da8c8780551f5995f7fc3042db29e7f50660ac2a1aa897eeed901

results/provisional_leaderboard.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd3593904840d62cd2e2c36e50c4654fcd29d70f3512ab56a3a71c1862985be7
+size 15354