Upload README.md with huggingface_hub

README.md

@@ -14,16 +14,6 @@ tags:
 - grammar
 - acceptability
 configs:
-- config_name: both
-  data_files:
-  - split: train
-    path: both/train-*
-  - split: validation
-    path: both/validation-*
-  - split: test
-    path: both/test-*
-  - split: test_unseen_lemmas
-    path: both/test_unseen_lemmas-*
 - config_name: en
   data_files:
   - split: train
@@ -44,6 +34,16 @@ configs:
     path: it/test-*
   - split: test_unseen_lemmas
     path: it/test_unseen_lemmas-*
+- config_name: both
+  data_files:
+  - split: train
+    path: both/train-*
+  - split: validation
+    path: both/validation-*
+  - split: test
+    path: both/test-*
+  - split: test_unseen_lemmas
+    path: both/test_unseen_lemmas-*
 - config_name: parallel
   data_files:
   - split: train
@@ -54,453 +54,74 @@ configs:
     path: parallel/test-*
   - split: test_unseen_lemmas
     path: parallel/test_unseen_lemmas-*
-dataset_info:
-- config_name: both
-  features:
-  - name: id
-    dtype: string
-  - name: lang
-    dtype: string
-  - name: text
-    dtype: string
-  - name: tokens
-    sequence: string
-  - name: clean_text
-    dtype: string
-  - name: clean_text_raw
-    dtype: string
-  - name: binary
-    dtype:
-      class_label:
-        names:
-          '0': ok
-          '1': not_ok
-  - name: multiclass
-    dtype:
-      class_label:
-        names:
-          '0': none
-          '1': word_order
-          '2': missing_word
-          '3': repeated_word
-          '4': agreement
-          '5': wrong_function_word
-          '6': multiple
-  - name: multilabel
-    sequence: int8
-    length: 5
-  - name: token_tags
-    sequence:
-      class_label:
-        names:
-          '0': O
-          '1': word_order
-          '2': missing_word
-          '3': repeated_word
-          '4': agreement
-          '5': wrong_function_word
-  - name: token_binary
-    sequence: int8
-  - name: n_corruptions
-    dtype: int8
-  - name: source
-    dtype: string
-  - name: source_id
-    dtype: string
-  - name: source_author
-    dtype: string
-  - name: license
-    dtype: string
-  - name: split
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 9298419
-    num_examples: 23865
-  - name: validation
-    num_bytes: 1179378
-    num_examples: 2983
-  - name: test
-    num_bytes: 1162240
-    num_examples: 2983
-  - name: test_unseen_lemmas
-    num_bytes: 1421536
-    num_examples: 3000
-  download_size: 4892903
-  dataset_size: 13061573
-- config_name: en
-  features:
-  - name: id
-    dtype: string
-  - name: lang
-    dtype: string
-  - name: text
-    dtype: string
-  - name: tokens
-    sequence: string
-  - name: clean_text
-    dtype: string
-  - name: clean_text_raw
-    dtype: string
-  - name: binary
-    dtype:
-      class_label:
-        names:
-          '0': ok
-          '1': not_ok
-  - name: multiclass
-    dtype:
-      class_label:
-        names:
-          '0': none
-          '1': word_order
-          '2': missing_word
-          '3': repeated_word
-          '4': agreement
-          '5': wrong_function_word
-          '6': multiple
-  - name: multilabel
-    sequence: int8
-    length: 5
-  - name: token_tags
-    sequence:
-      class_label:
-        names:
-          '0': O
-          '1': word_order
-          '2': missing_word
-          '3': repeated_word
-          '4': agreement
-          '5': wrong_function_word
-  - name: token_binary
-    sequence: int8
-  - name: n_corruptions
-    dtype: int8
-  - name: source
-    dtype: string
-  - name: source_id
-    dtype: string
-  - name: source_author
-    dtype: string
-  - name: license
-    dtype: string
-  - name: split
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 4875568
-    num_examples: 11865
-  - name: validation
-    num_bytes: 620158
-    num_examples: 1483
-  - name: test
-    num_bytes: 612619
-    num_examples: 1483
-  - name: test_unseen_lemmas
-    num_bytes: 742534
-    num_examples: 1500
-  download_size: 2590903
-  dataset_size: 6850879
-- config_name: it
-  features:
-  - name: id
-    dtype: string
-  - name: lang
-    dtype: string
-  - name: text
-    dtype: string
-  - name: tokens
-    sequence: string
-  - name: clean_text
-    dtype: string
-  - name: clean_text_raw
-    dtype: string
-  - name: binary
-    dtype:
-      class_label:
-        names:
-          '0': ok
-          '1': not_ok
-  - name: multiclass
-    dtype:
-      class_label:
-        names:
-          '0': none
-          '1': word_order
-          '2': missing_word
-          '3': repeated_word
-          '4': agreement
-          '5': wrong_function_word
-          '6': multiple
-  - name: multilabel
-    sequence: int8
-    length: 5
-  - name: token_tags
-    sequence:
-      class_label:
-        names:
-          '0': O
-          '1': word_order
-          '2': missing_word
-          '3': repeated_word
-          '4': agreement
-          '5': wrong_function_word
-  - name: token_binary
-    sequence: int8
-  - name: n_corruptions
-    dtype: int8
-  - name: source
-    dtype: string
-  - name: source_id
-    dtype: string
-  - name: source_author
-    dtype: string
-  - name: license
-    dtype: string
-  - name: split
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 4422851
-    num_examples: 12000
-  - name: validation
-    num_bytes: 559220
-    num_examples: 1500
-  - name: test
-    num_bytes: 549621
-    num_examples: 1500
-  - name: test_unseen_lemmas
-    num_bytes: 679002
-    num_examples: 1500
-  download_size: 2325203
-  dataset_size: 6210694
-- config_name: parallel
-  features:
-  - name: id
-    dtype: string
-  - name: lang
-    dtype: string
-  - name: text
-    dtype: string
-  - name: tokens
-    sequence: string
-  - name: clean_text
-    dtype: string
-  - name: clean_text_raw
-    dtype: string
-  - name: binary
-    dtype:
-      class_label:
-        names:
-          '0': ok
-          '1': not_ok
-  - name: multiclass
-    dtype:
-      class_label:
-        names:
-          '0': none
-          '1': word_order
-          '2': missing_word
-          '3': repeated_word
-          '4': agreement
-          '5': wrong_function_word
-          '6': multiple
-  - name: multilabel
-    sequence: int8
-    length: 5
-  - name: token_tags
-    sequence:
-      class_label:
-        names:
-          '0': O
-          '1': word_order
-          '2': missing_word
-          '3': repeated_word
-          '4': agreement
-          '5': wrong_function_word
-  - name: token_binary
-    sequence: int8
-  - name: n_corruptions
-    dtype: int8
-  - name: source
-    dtype: string
-  - name: source_id
-    dtype: string
-  - name: source_author
-    dtype: string
-  - name: license
-    dtype: string
-  - name: split
-    dtype: string
-  - name: pair_id
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 44740
-    num_examples: 120
-  - name: validation
-    num_bytes: 7170
-    num_examples: 20
-  - name: test
-    num_bytes: 2451
-    num_examples: 6
-  - name: test_unseen_lemmas
-    num_bytes: 1444
-    num_examples: 4
-  download_size: 60852
-  dataset_size: 55805
 ---
 
 # Lapsus
 
-## Purpose
+Lapsus is a small dataset of grammatical and ungrammatical sentences in English and Italian. It was made for the course **Elaborazione del Linguaggio Naturale** at [Kore University of Enna](https://www.uke.it), for hands-on exercises with BERT-style models. The code that builds it is on [GitHub](https://github.com/MorenoLaQuatra/lapsus).
 
-Lapsus is a bilingual (English and Italian) grammatical acceptability dataset for teaching NLP with transformers (BERT, RoBERTa, XLM-R). Clean sentences come from open corpora, pass a chain of filters, and half of them are corrupted with named operations whose output is verified by independent checks. The same rows support four tasks: binary acceptability, multiclass error type, multi-label error types and token-level error tagging.
+## How it works
 
-Canonical releases live on GitHub (https://github.com/MorenoLaQuatra/lapsus/releases); this Hub repository (morenolq/lapsus) is a mirror. Dataset version: 0.1.0, seed 42.
+We take clean sentences from [Tatoeba](https://tatoeba.org) and break some of them on purpose, keeping track of what we broke and where. There are five kinds of errors:
 
-## Configs and splits
+- `word_order`: two nearby words are swapped (*"i never your cut hair"*)
+- `missing_word`: a function word is deleted (*"she went the station"*)
+- `repeated_word`: a word is doubled (*"he is is tired"*)
+- `agreement`: a form no longer agrees (*"the dogs runs"*, *"la ragazzi"*)
+- `wrong_function_word`: the wrong preposition or article (*"in"* instead of *"on"*, *"di"* instead of *"da"*)
 
-| config | licence | train | validation | test | test_unseen_lemmas | test_ood_source | test_natural |
-|---|---|---|---|---|---|---|---|
-| en | apache-2.0 | 11865 | 1483 | 1483 | 1500 |  |  |
-| it | cc-by-2.0 | 12000 | 1500 | 1500 | 1500 |  |  |
-| both | cc-by-2.0 | 23865 | 2983 | 2983 | 3000 |  |  |
-| parallel | cc-by-2.0 | 120 | 20 | 6 | 4 |  |  |
+In every split 50% of the sentences are clean, 30% have one error and 20% have two or three (labelled `multiple`).
 
-- `both` concatenates `en` and `it` (use the `lang` column).
-- `parallel` holds Tatoeba-linked English/Italian translation pairs (`pair_id`), both members in the same split, for cross-lingual demos. Members can be clean or corrupted independently.
+Each broken sentence is checked before it is kept: the new word must be a real form, the result must not be another clean sentence, and a small language model must like it less than the original, that is $\log p(\text{clean}) - \log p(\text{corrupted}) \geq \delta$ (average per token). The checks are automatic, so a few mistakes are still in there.
 
-## Schema
+## Tasks
 
-One row per source sentence; no source sentence is reused.
+The same rows support four tasks. For *"the dogs runs fast"*:
 
-| column | type | description |
+| task | column | label |
 |---|---|---|
-| id | string | `<lang>-<source>-<source_id>` |
-| lang | string | `en` or `it` |
-| text | string | the (possibly corrupted) sentence after normalization |
-| tokens | list of string | tokens of `text` (see Tokenization) |
-| clean_text | string | the clean sentence, same normalization |
-| clean_text_raw | string | the clean sentence as in the source (NFC, trimmed) |
-| binary | ClassLabel | ok, not_ok |
-| multiclass | ClassLabel | none, word_order, missing_word, repeated_word, agreement, wrong_function_word, multiple |
-| multilabel | 5 x int8 | fixed order: word_order, missing_word, repeated_word, agreement, wrong_function_word |
-| token_tags | list of ClassLabel | O, word_order, missing_word, repeated_word, agreement, wrong_function_word |
-| token_binary | list of int8 | 1 if the token carries an error tag |
-| n_corruptions | int8 | 0, 1, 2 or 3 |
-| source, source_id, source_author | string | provenance (Tatoeba sentence id and contributor username) |
-| license | string | SPDX licence of the sentence |
-| split | string | split name |
+| binary | `binary` | `not_ok` |
+| error type | `multiclass` | `agreement` |
+| error types (multi-label) | `multilabel` | $[0, 0, 0, 1, 0]$ |
+| token tagging | `token_tags` | *runs* is tagged `agreement`, the rest `O` |
 
-In `test_natural` only `binary` and `token_binary` are defined (real errors have no operation type): `multiclass` and `token_tags` are -1 (null in csv/jsonl) and `multilabel`, `n_corruptions` are -1.
+The `multilabel` order is `word_order`, `missing_word`, `repeated_word`, `agreement`, `wrong_function_word`. When a word is missing, the token right after the gap gets the tag.
 
-CSV files store list columns as JSON strings and labels as names; JSONL stores labels as names; the Hub/DatasetDict version stores ClassLabel integers.
-
-Mix per split: 50 percent clean, 30 percent one corruption (balanced over the 5 operations), 20 percent two or three different corruptions.
-
-## Corruption operations
-
-- **word_order**: swap two tokens at distance up to 3; both tokens are tagged. Swaps must involve a determiner, preposition, auxiliary, pronoun or verb; adverbs never move; English never swaps a subject with an auxiliary (that makes a valid question); Italian never swaps ADJ and NOUN, two NOUNs, two arguments of one verb, or a subject with its verb.
-- **missing_word**: delete one function word (article, preposition, non-modal auxiliary, infinitival "to"; Italian articles, simple and articulated prepositions, auxiliaries and reflexive clitics).
-- **repeated_word**: duplicate a token right after itself; the duplicate is tagged. Words whose doubling can be grammatical ("that that", "had had", "molto molto") are excluded.
-- **agreement**: EN subject-verb (LemmInflect VBZ/VBP, be, negative contractions), determiner-noun number (this/these, a + plural noun), a/an. IT gender and number across DET (articles, articulated prepositions, possessives, demonstratives), NOUN and ADJ; subject-verb person and number; predicative adjective and participle (with essere) agreement. Forms come from Morph-it! (same lemma, different feature) and must be real word forms that cannot also carry the original feature value. Only relations present in the clean parse (nsubj, det, amod) are targeted.
-- **wrong_function_word**: confusion sets. EN: in/on/at, to/for/of, and a restricted a/the confusion ("a" for "the" only before plural nouns or superlatives). IT: a/di/da/in/su/per/con, articulated prepositions (preposition part changed), article allomorphs il/lo, i/gli, un/uno (article gender/number changes belong to agreement).
-- **multiple**: 2 or 3 different operations applied sequentially on non-overlapping tokens.
-
-## Missing-word tag convention
-
-A deleted word leaves no token to tag, so the token right after the gap is tagged `missing_word`; if the gap is at the end of the sentence, the last token is tagged.
-
-## Tokenization and normalization
-
-Tokens are whitespace separated. In Italian, elided forms are split off and keep their apostrophe ("l'uomo" -> "l'", "uomo"; "dell'acqua" -> "dell'", "acqua") and `text` glues them back. Punctuation stays attached to words. Normalization: `lower_nopunct` (lowercase and strip the final punctuation by default); `clean_text_raw` keeps the original form.
-
-## Sources and attribution
-
-- **tatoeba** (https://downloads.tatoeba.org/exports/per_language/), licence CC-BY-2.0, export date(s): Sat, 31 Oct 2026 06:39:48 GMT, Sat, 31 Oct 2026 06:41:23 GMT, Sat, 31 Oct 2026 06:52:53 GMT. Sentences from Tatoeba (https://tatoeba.org), by the Tatoeba contributors named in the source_author column. Licensed CC BY 2.0 FR unless marked CC0 1.0.
-
-Full attribution, including every Tatoeba contributor, is in ATTRIBUTION.md.
-
-## Licence mode and pool sizes
-
-Requested mode: `auto`. Labels, corruptions and annotations are Apache-2.0.
-
-| lang | mode used | reason | permissive pool | total pool | threshold | config licence |
-|---|---|---|---|---|---|---|
-| en | apache-strict | auto: permissive pool 16632 >= threshold 15000 | 16632 | 30492 | 15000 | apache-2.0 |
-| it | source-licenses | auto: permissive pool 7 < threshold 15000, falling back to source licences | 7 | 30890 | 15000 | cc-by-2.0 |
-
-Hub repository licence (most restrictive across configs): `cc-by-2.0`.
-
-
-## Filters and verification
-
-Clean-sentence filters: (a) NFC, one sentence, 5 to 20 tokens; (b) no digits, URLs, emoji, quotes, brackets, symbols, all-caps, at most one proper noun, any single proper noun capped at 2 percent of the pool; (c) language ID (lingua); (d) LanguageTool with no grammar or typo match (style, casing and punctuation rules ignored); (e) Stanza parse with exactly one root and no X tags; (f) causal LM filter dropping the lowest 5 percent mean log-prob per length bucket; (g) exact, normalized and MinHash near-duplicate removal on lemma shingles.
-
-- en rejections: a_normalize_length 376302, b_lexical 360959, c_langid 390, d_languagetool 1128, e_parse 113, f_lm 1839, g_dedup_exact 0, g_dedup_normalized 3, g_dedup_near 271, b2_name_cap 4208
-- it rejections: a_normalize_length 300398, b_lexical 80147, c_langid 745, d_languagetool 1128, e_parse 35, f_lm 2004, g_dedup_exact 0, g_dedup_normalized 169, g_dedup_near 3428, b2_name_cap 3609
-
-Verification of corrupted rows: differs from the clean sentence; equals no clean-pool sentence; agreement re-checked against the lexicon and the clean parse; tags aligned and label views consistent; causal-LM minimal pair mean_logprob(clean) - mean_logprob(corrupted) >= delta. LanguageTool flag rates are diagnostic only.
+## Configs and splits
 
-| lang | op | acceptance rate | LM diff (clean minus corrupted) | LanguageTool flag rate |
+| config | `train` | `validation` | `test` | `test_unseen_lemmas` |
 |---|---|---|---|---|
-| en | delta | 0.2931 | | |
-| en | agreement | 0.8463 | mean 0.6704, p05 0.2088, p50 0.6329, p95 1.2737 (n=1501) | 0.6663 |
-| en | missing_word | 0.9432 | mean 1.1297, p05 0.334, p50 1.0809, p95 2.1028 (n=1501) | 0.098 |
-| en | multiple | 0.9924 | mean 2.2748, p05 0.7465, p50 2.2206, p95 4.0155 (n=4963) | 0.7964 |
-| en | repeated_word | 0.7986 | mean 0.5242, p05 0.1246, p50 0.5014, p95 1.0081 (n=1033) | 0.9705 |
-| en | word_order | 0.9879 | mean 2.1046, p05 0.7121, p50 2.0346, p95 3.748 (n=1501) | 0.2878 |
-| en | wrong_function_word | 0.8187 | mean 0.7581, p05 0.1189, p50 0.6938, p95 1.6234 (n=1501) | 0.1082 |
-| it | delta | 0.0 | | |
-| it | agreement | 0.8042 | mean 0.4275, p05 -0.1925, p50 0.423, p95 1.0904 (n=1517) | 0.0455 |
-| it | missing_word | 0.8075 | mean 0.7946, p05 -0.038, p50 0.7511, p95 1.7632 (n=1517) | 0.001 |
-| it | multiple | 0.8154 | mean 1.1644, p05 -0.0937, p50 1.103, p95 2.6638 (n=5014) | 0.5776 |
-| it | repeated_word | 0.6762 | mean 0.1622, p05 -0.4701, p50 0.1878, p95 0.7297 (n=1517) | 0.998 |
-| it | word_order | 0.8433 | mean 1.1832, p05 0.0676, p50 1.1512, p95 2.4112 (n=1517) | 0.0545 |
-| it | wrong_function_word | 0.8639 | mean 0.4446, p05 -0.1298, p50 0.4316, p95 1.129 (n=1517) | 0.0081 |
-
-Human audit: audit/clean_{lang}.csv (200 clean rows) and audit/corrupted_{lang}.csv (200 corrupted rows, stratified by type) are provided for annotation; results are not filled in yet.
+| `en` | 11865 | 1483 | 1483 | 1500 |
+| `it` | 12000 | 1500 | 1500 | 1500 |
+| `both` | 23865 | 2983 | 2983 | 3000 |
+| `parallel` | 120 | 20 | 6 | 4 |
 
-## Splits
+`both` is `en` plus `it` with a `lang` column. `parallel` is a small set of English-Italian translation pairs (`pair_id`). In `test_unseen_lemmas` the content words never appear in the other splits, so it tests generalization rather than memorization.
 
-Splits are built per language and grouped by source sentence, with automated leakage tests. `test_unseen_lemmas` uses reserved content lemmas that never occur in train, validation or test. `test_ood_source` uses a different source when enabled and allowed by the licence mode. `test_natural` contains real learner errors (CoEdIT src vs tgt aligned with difflib for English; Italian only from user-supplied files). The `parallel` config keeps translation pairs in the same split, and Italian translations of English sentences are forced into the English sentence's split.
+```python
+from datasets import load_dataset
 
-- en: 335 reserved lemmas; dropped: {"unused": 1}
-- it: 329 reserved lemmas; dropped: {"forced_unseen_without_reserved_lemma": 9, "unused": 11081}
+ds = load_dataset("morenolq/lapsus", "it")
+```
 
-## Tool and model licences
+Other useful columns: `tokens`, `clean_text` (the original sentence), `n_corruptions`, `source_id`, `source_author` and `license`. Text is lowercased and the final punctuation is removed; Italian elisions are split (*l'uomo* becomes *l'* + *uomo*).
 
-| tool or model | version | licence | notes |
-|---|---|---|---|
-| Stanza | 1.15.0 | Apache-2.0 | Library code. |
-| Stanza English models (default "combined" package) |  | CC-BY-SA-4.0 (EWT) plus other UD English treebanks | Trained on UD English treebanks; the combined package mixes EWT, GUM and others. Check each treebank licence before redistributing parses. |
-| Stanza Italian models (default "combined" package) |  | CC-BY-NC-SA (ISDT, VIT, PoSTWITA, TWITTIRO training data) **NON-COMMERCIAL** | Model weights derive from non-commercial treebanks. Used only as a tool; no parses are redistributed. |
-| LanguageTool | 6.6 | LGPL-2.1-or-later |  |
-| Morph-it! | 0.48 | CC-BY-SA-2.0 OR LGPL-3.0 | Used as a lexicon to generate and validate Italian word forms; not redistributed. |
-| LemmInflect | 0.2.3 | MIT |  |
-| lingua-language-detector | 2.1.1 | Apache-2.0 |  |
-| datasketch | 2.0.0 | MIT |  |
-| openai-community/gpt2 |  | MIT | English causal LM for the LM filter and the minimal-pair verifier. |
-| Qwen/Qwen2.5-0.5B |  | Apache-2.0 | Italian causal LM for the LM filter and the verifier (multilingual model). sapienzanlp/Minerva-350M-base-v1.0 (Apache-2.0, Italian-specific, gated) is a drop-in alternative via lm.it in the config. |
+## Licence
 
-Tool outputs (parses, LM scores) are not redistributed; they only drive filtering, target selection and verification.
-
-## Related work
-
-- **BLiMP**: Warstadt et al. (2020), The Benchmark of Linguistic Minimal Pairs for English, TACL.
-- **CoLA**: Warstadt, Singh and Bowman (2019), Neural Network Acceptability Judgments, TACL.
-- **ItaCoLA**: Trotta et al. (2021), Monolingual and Cross-Lingual Acceptability Judgments with the Italian CoLA corpus, Findings of EMNLP.
-- **BLiMP-IT**: Italian minimal-pair benchmarks modelled on BLiMP (see the BLiMP-IT and related Italian minimal-pair releases).
-- **AcCompl-it**: Brunato et al. (2020), AcCompl-it at EVALITA 2020: acceptability and complexity prediction for Italian.
-- **DaLA**: Danish linguistic acceptability dataset built by corrupting correct sentences with controlled error types.
+The dataset is released as `cc-by-2.0` because it follows its Tatoeba sources: English sentences are CC0 (public domain), Italian sentences are CC BY 2.0 FR. Each row has its own `license` column and every contributor is listed in `ATTRIBUTION.md`. Labels and code are Apache-2.0.
 
 ## Limitations
 
-- Corruptions are synthetic and rule based: they cover five error families, not the full range of learner errors (see test_natural for real errors).
-- Verification is automatic. A corrupted sentence can occasionally still be grammatical (for example a preposition swap that yields another valid reading), and filters can let through awkward clean sentences. The audit files estimate both error rates.
-- Tatoeba sentences are short, conversational and dominated by a few contributors and recurring names; the proper-noun cap reduces but does not remove this.
-- Default normalization lowercases text and strips final punctuation, so capitalization and punctuation cues are not available to models.
-- Surface cues: repeated words and some word-order swaps are detectable with shallow features; see the TF-IDF baselines before claiming a model learned syntax.
-- The Italian LanguageTool grammar coverage is limited, so the Italian clean filter relies more on the parser and LM filters.
+- The errors are synthetic and cover only five families, not everything learners do.
+- Tatoeba sentences are short and conversational.
+- Some errors (like repeated words) are easy to spot with shallow features, so try a TF-IDF baseline before claiming a model learned syntax.
+
+## Citation
 
+```bibtex
+@misc{lapsus,
+  title  = {Lapsus: a bilingual grammatical acceptability dataset for teaching NLP},
+  howpublished = {\url{https://github.com/MorenoLaQuatra/lapsus}},
+  year   = {2026}
+}
+```