Add frequency-preserving surname phonetic G2P evaluation dataset

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+rows.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,52 @@
+---
+license: other
+task_categories:
+- text-generation
+tags:
+- g2p
+- pronunciation
+- names
+- evaluation
+configs:
+- config_name: default
+  data_files:
+  - split: eval
+    path: rows.jsonl
+- config_name: grouped
+  data_files:
+  - split: train
+    path: groups.jsonl
+- config_name: audit
+  data_files:
+  - split: train
+    path: excluded.jsonl
+---
+
+# NetJets surname phonetic G2P evaluation
+
+Source: `Namecoach/netjets-name-asr`, Hub revision `dd9ca45ff76c1cf5017d03bbd3558f625420928f`, train split.
+
+This benchmark presents a written surname grapheme and expects one of the observed custom written phonetic spellings for that surname. The phonetic strings are retained exactly after outer whitespace trimming; capitalization encodes stress and hyphens/parentheses are meaningful notation.
+
+## Files
+
+- `rows.jsonl`: one row per retained source occurrence. Do not deduplicate this file: its repeated rows preserve the source pronunciation frequency.
+- `groups.jsonl`: one row per grapheme with its observed variant counts and probabilities. Use this for distributional evaluation and allowed-output validation.
+- `excluded.jsonl`: source rows that could not be mapped to one auditable grapheme token, with the reason and candidate scores.
+- `manifest.json`: source hash, counts, extraction policy, and frequency policy.
+
+## Row contract
+
+Each row contains `grapheme`, `phonetic`, a stable source `id`, source provenance, and alignment metadata. A prediction is valid only if it exactly equals one of the `phonetic` variants listed for that row's `grapheme` in `groups.jsonl`.
+
+## Alignment policy
+
+The dataset's `name` field is an ASR/spoken reconstruction and can contain extra or missing tokens, so the final whitespace token is not always the labeled surname. The builder uses a documented token-to-phonetic character-similarity heuristic with a surname prior, removes common suffixes such as Jr/III, and excludes ambiguous or unalignable rows instead of fabricating grapheme labels.
+
+## Counts
+
+- Source rows: 36,704
+- Retained eval occurrences: 35,891
+- Unique graphemes: 15,938
+- Unique grapheme/phonetic pairs: 27,644
+- Excluded rows: 813

manifest.json

@@ -0,0 +1,21 @@
+{
+  "dataset": "Namecoach/netjets-name-asr",
+  "dataset_revision": "dd9ca45ff76c1cf5017d03bbd3558f625420928f",
+  "eval_rows": 35891,
+  "excluded_policy": "exclude missing fields, ambiguous token alignment, and unalignable token alignment; retain full details in excluded.jsonl",
+  "excluded_rows": 813,
+  "files": {
+    "excluded": "excluded.jsonl",
+    "groups": "groups.jsonl",
+    "rows": "rows.jsonl"
+  },
+  "frequency_policy": "one eval row per retained source occurrence; no deduplication or reweighting",
+  "grapheme_policy": "auditable token alignment; surname-prior when plausible, otherwise unique token similarity; suffixes are removed",
+  "phonetic_policy": "phonetic retained exactly after outer whitespace trimming; capitalization, hyphens, parentheses, and stress notation preserved",
+  "source_core_name_phonetic_sha256": "1b7c6b006d26a966605485d1865d07096baa1118cb3c4b85caf899d209ac6c4d",
+  "source_local_mirror": "/home/rezo/work/netjets_name_asr_pro/data",
+  "source_rows": 36704,
+  "source_split": "train",
+  "unique_grapheme_phonetic_pairs": 27644,
+  "unique_graphemes": 15938
+}

rows.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6eea96de28ad321c9ca50f7c4f4c2290b05495c43750154c40581421a5cec29
+size 15071353