Upload folder using huggingface_hub

README.md

@@ -49,7 +49,7 @@ configs:
 A morphological inflection and analysis benchmark for Kabyle (Taqbaylit, `kab`, Latin script), from the
 [AƔBALU](https://huggingface.co/agbalu) project.
 
-**344,745 inflected verb form entries** across **13,226 unique verb lemmas**, partitioned into
+**336,151 inflected verb form entries** across **13,226 unique verb lemmas**, partitioned into
 paradigmatically sealed splits (0 paradigm leakage), plus **6,198 complete verb conjugation tables**.
 
 ```python
@@ -64,14 +64,14 @@ paradigms = load_dataset("agbalu/KabInflect", "paradigms", split="train")
 
 Morphological Inflection (Seq2Seq Morphological Generation): `(lemma, feats)` → `form`.
 
-344,745 total entries across 13,226 verb lemmas.
+336,151 total entries across 13,226 verb lemmas.
 
 | split | lemmas | inflected forms |
 |---|---|---|
-| `train` | 10,580 | 276,717 |
-| `dev` | 1,323 | 34,109 |
-| `test` | 1,323 | 33,919 |
-| **total** | **13,226** | **344,745** |
+| `train` | 10,580 | 270,026 |
+| `dev` | 1,323 | 33,060 |
+| `test` | 1,323 | 33,065 |
+| **total** | **13,226** | **336,151** |
 
 ### Schema
 
@@ -82,13 +82,13 @@ Morphological Inflection (Seq2Seq Morphological Generation): `(lemma, feats)`
 | `feats` | string | Universal Dependencies (UD) FEATS (e.g. `Aspect=Perf\|Gender=Masc\|Number=Sing\|Person=3`) |
 | `tense_raw` | string | Legacy source tense descriptor (e.g. `prétérit`, `aoriste intensif`) |
 | `person_raw` | string | Legacy source person descriptor (e.g. `3s_m`, `1p`) |
-| `form` | string | Inflected surface form in canonical Kabyle orthography |
+| `form` | string | Inflected surface form, normalised; see Orthography — emphatic `ṛ` is unmarked |
 
 ## `analysis`
 
 Morphological Analysis and Lemmatisation: `form` → `(lemma, feats)`.
 
-Same 344,745 entries, indexed by inflected surface form.
+Same 336,151 entries, indexed by inflected surface form.
 
 ## `paradigms`
 
@@ -101,19 +101,35 @@ Same 344,745 entries, indexed by inflected surface form.
 `agbalu/KabInflect` fixes this:
 - Entries are grouped strictly by **verb lemma** (`infinitif`).
 - Splitting is performed at the **lemma level** with seed 42.
-- **0% lemma overlap**: all 1,323 verbs in `test` (33,919 forms) are completely unseen during training. A system must learn the underlying rules of Kabyle morphology to succeed.
+- **0% lemma overlap**: all 1,323 verbs in `test` (33,065 forms) are completely unseen during training. A system must learn the underlying rules of Kabyle morphology to succeed.
 
 ## Orthography
 
-All Kabyle text is normalised under normaliser `1.3.0+rules1.0.0` (81 rules), preserving sub-dot emphatic distinctions (`ɣ ḍ ḥ ṭ ṛ ẓ ṣ`). The source text carries 0.00% homoglyph error.
+All Kabyle text is normalised under normaliser `1.3.0+rules1.0.0` (81 rules). The source
+text carries 0.00% homoglyph error and 0.00% legacy-font mojibake.
+
+🔴 **The emphatic `ṛ` is not represented, and this is a property of the source.** Measured
+over all 270,026 training forms, `ṛ` U+1E5B occurs **0 times**, while `ḍ` appears in 14.43%
+of forms, `ḥ` in 12.14%, `ṣ` in 3.36%, `ẓ` in 2.73% and `ṭ` in 4.37%. The upstream
+conjugation resource writes plain `r` throughout — its lemmas are `ruḥ`, `ṣber` and `ɛreḍ`,
+not `ṛuḥ`, `ṣbeṛ` and `ɛṛeḍ`.
+
+The normaliser did not remove it; there was never anything to preserve. **The `ṛ`/`r`
+contrast is therefore neutralised in this dataset**, and it is *not* neutralised in
+[`agbalu/KabLex`](https://huggingface.co/datasets/agbalu/KabLex), which carries both `ṛuḥ`
+and `ruḥ`. A model trained here will not produce emphatic `ṛ`, and scoring its output
+against text that marks it will count every such form wrong.
+
+Restoring it would mean deciding, for each `r`, whether the verb takes the emphatic — a
+lexical question this dataset has no evidence for. It is disclosed rather than guessed.
 
 ## Evaluation Protocol & Baseline Floor
 
-The benchmark evaluates morphological generation (`inflection`) and parsing (`analysis`) on **1,323 held-out test verb lemmas (33,919 unseen test forms)**.
+The benchmark evaluates morphological generation (`inflection`) and parsing (`analysis`) on **1,323 held-out test verb lemmas (33,065 unseen test forms)**.
 
 | task | system | exact match | CER |
 |---|---|---|---|
-| inflection | copy the lemma unchanged | 4.11% | 30.69% |
+| inflection | copy the lemma unchanged | 4.07% | 30.73% |
 
 ```bash
 make bench TASK=inflect
@@ -121,7 +137,7 @@ make bench TASK=inflect
 
 **The floor is not zero, and that is why it is measured rather than assumed.** Kabyle's
 imperative singular *is* the citation form for most verbs, so a copy is correct in one cell
-of nearly every paradigm — 4.11% of 33,919 test forms. A headline exact match read against
+of nearly every paradigm — 4.07% of 33,065 test forms. A headline exact match read against
 zero would be flattered by exactly that much.
 
 No system has yet been scored on the `analysis` direction.
@@ -131,6 +147,12 @@ paradigms: no test verb appears in any training cell.
 
 ## Known Limits
 
+- **No emphatic `ṛ`.** See Orthography above — 0 occurrences in 270,026 forms, because the
+  source resource does not mark it. This is the most consequential limit here.
+- **Deduplicated.** The raw shards repeat rows verbatim: one verb's intensive stems can
+  generate the same cell twice. **8,594 identical rows were dropped** (344,745 → 336,151),
+  so no form is silently reweighted during training. Earlier revisions of this dataset
+  carried them.
 - **Verb-focused.** Covers Kabyle verbal morphology (the largest morphological paradigm family in the language); nominal annexed-state paradigms are covered separately in `agbalu/KabLex`.
 - **Gloss language.** Paradigm translations are French, reflecting the source dictionary.
 - **Standard orthography.** Uses standard Kabyle Latin script; does not cover dialectal spellings or Tifinagh (see `agbalu/KabTifinagh` for script conversion).

analysis/dev.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:18f8150c115b1747e2857526c6cbb28ee679f0b750c81b63a712ad4a1f88c760
-size 492588
+oid sha256:c4d820de47f0b63c40fef61eb9a5d0bee95f0f35f39bcb4a97acb660d7ea832b
+size 481478

analysis/test.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a8c39f7c699be2240fdacd863c38ab3326fe1d2fb873744fb61041ab7b88e1a1
-size 491641
+oid sha256:a84bc835369e647d8a39f35495ddc0c31576fba2eb6eea7c20e7468c1fbc4198
+size 482544

analysis/train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6d1458de4cc7f420df91aa81dccdc9541bcf6a8bb9065c9e5af441ce226c709f
-size 3452484
+oid sha256:74edcf3b942233b536cea18e0119a599531e479f3614d04be1cb73f31b0c873f
+size 3388634

dataset.stats.json

@@ -4,44 +4,44 @@
     {
       "config": "inflection",
       "split": "train",
-      "rows": 276717,
+      "rows": 270026,
       "path": "inflection/train.parquet",
-      "bytes": 3459886
+      "bytes": 3396080
     },
     {
       "config": "inflection",
       "split": "dev",
-      "rows": 34109,
+      "rows": 33060,
       "path": "inflection/dev.parquet",
-      "bytes": 495071
+      "bytes": 483911
     },
     {
       "config": "inflection",
       "split": "test",
-      "rows": 33919,
+      "rows": 33065,
       "path": "inflection/test.parquet",
-      "bytes": 490734
+      "bytes": 481761
     },
     {
       "config": "analysis",
       "split": "train",
-      "rows": 276717,
+      "rows": 270026,
       "path": "analysis/train.parquet",
-      "bytes": 3452484
+      "bytes": 3388634
     },
     {
       "config": "analysis",
       "split": "dev",
-      "rows": 34109,
+      "rows": 33060,
       "path": "analysis/dev.parquet",
-      "bytes": 492588
+      "bytes": 481478
     },
     {
       "config": "analysis",
       "split": "test",
-      "rows": 33919,
+      "rows": 33065,
       "path": "analysis/test.parquet",
-      "bytes": 491641
+      "bytes": 482544
     },
     {
       "config": "paradigms",
@@ -51,5 +51,5 @@
       "bytes": 2456801
     }
   ],
-  "rows": 695688
+  "rows": 678500
 }

inflection/dev.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0342069072c23a9a60ac6fd6966342c0a1c448eafe23df6b2df2cba7b236c084
-size 495071
+oid sha256:ac765b990ca6619ddebb6341ac1ecc602af8bff3bcef7017f6e9b9b11ede0f60
+size 483911

inflection/test.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:105fe6ac45491c0263475dac09d68198104bb0b3afdf8eb76eb854b6eafec1a3
-size 490734
+oid sha256:b634885429788eb863a2eb7c8c23e04b37341efa0192e26ff792763d0a4f8ee8
+size 481761

inflection/train.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:76118afb5eaa7c648d5a66aece48477e741c41766fba38d33eff00bb6134ded6
-size 3459886
+oid sha256:aa46ab1191c7ef0d71dcb1cba103edc1f357e83e6b1b94deab389efaf423fbbf
+size 3396080