Update README: gold corpus, remove GMB refs, fix folder structure

README.md

@@ -18,6 +18,7 @@ tags:
   - conllu
   - spacy
   - gpt-validated
+  - llm-gold-filtered
   - knowledge-distillation
 size_categories:
   - 100K<n<1M
@@ -27,45 +28,50 @@ size_categories:
 
 Multi-domain NLP corpus for English with annotations for NER (18 entity types), POS tagging (17 UPOS tags), dependency parsing (dep2label), and sentence classification (9 dialog act types). Built for training [kniv](https://github.com/rustic-ai/kniv-nlp-models) multi-task NLP models for the [uniko](https://github.com/rustic-ai/uniko) cognitive memory system.
 
-All data is commercially licensed (CC BY-SA 4.0 compatible). No CoNLL-2003 or other research-only datasets.
+All data is commercially licensed (CC BY-SA 4.0 compatible). No CoNLL-2003, GMB, or other research-only datasets.
 
 ## Repository Structure
 
 ```
 dragonscale-ai/kniv-corpus-en/
-├── data/                           # HF auto-detected Parquet (for load_dataset)
-│   ├── train-00000-of-00001.parquet    628,577 sentences
-│   ├── dev-00000-of-00001.parquet       65,731 sentences
-│   └── test-00000-of-00001.parquet      65,732 sentences
-├── train.conllu                    # CoNLL-U format (raw linguistic annotation)
-├── dev.conllu
-├── test.conllu
-├── metadata.json                   # Split sizes, domain list
-├── prepared/                       # Pre-processed training-ready JSON splits
-│   ├── deberta-v3-large-nlp-en/        Standard prepared data
-│   │   ├── ner_train.json              NER training examples (45K subsampled)
-│   │   ├── ner_dev.json                NER dev examples
-│   │   ├── ner_test.json               NER test examples
-│   │   ├── ud_train.json               POS + Dep training (UD EWT, 12K)
-│   │   ├── ud_dev.json                 POS + Dep dev
-│   │   ├── ud_test.json                POS + Dep test
-│   │   └── label_vocabs.json           All label vocabularies
-│   └── deberta-v3-large-nlp-en-gold/   LLM-validated gold data (recommended)
-│       ├── ner_train.json              NER gold (36K, Qwen3-8B validated)
-│       ├── ner_dev.json                NER gold dev (4.5K)
-│       ├── ner_test.json               NER gold test (4.5K)
-│       ├── ud_train.json               POS gold (10.3K, Qwen3-8B validated)
-│       ├── ud_dev.json                 POS dev
-│       ├── ud_test.json                POS test
-│       └── label_vocabs.json           All label vocabularies
+├── data/                               # Full corpus Parquet (for load_dataset)
+│   ├── train-00000-of-00001.parquet        525,850 sentences (127.7 MB)
+│   ├── dev-00000-of-00001.parquet           65,731 sentences (13.8 MB)
+│   └── test-00000-of-00001.parquet          65,732 sentences (14.0 MB)
+├── corpus/
+│   └── gold/                           # LLM-validated gold corpus (recommended)
+│       ├── train.parquet                   237,646 sentences (56.8 MB)
+│       ├── dev.parquet                      65,731 sentences (13.8 MB)
+│       ├── test.parquet                     65,732 sentences (14.0 MB)
+│       └── progress.json                   Filter progress metadata
+├── train.conllu                        # Full corpus in CoNLL-U format (547.7 MB)
+├── dev.conllu                          # (67.9 MB)
+├── test.conllu                         # (68.8 MB)
+├── metadata.json                       # Split sizes, domain list
 └── README.md
 ```
 
+### Full corpus vs Gold corpus
+
+| | Full (`data/`) | Gold (`corpus/gold/`) |
+|---|---|---|
+| **Train sentences** | 525,850 | 237,646 |
+| **Validation** | spaCy + GPT corrections | + Qwen3-8B per-example validation |
+| **Checks applied** | None | NER entities, CLS labels, POS tags |
+| **Recommended for** | Exploration, analysis | Model training |
+
+The gold corpus was produced by validating every sentence with Qwen3-8B (64 concurrent requests via vLLM on A100). Each sentence passed three checks:
+- **NER**: every entity span validated ("Is '{entity}' correctly tagged as {type}?")
+- **CLS**: dialog act label validated with conversational context
+- **POS**: up to 8 POS tags per sentence validated against Universal POS tagset
+
+Sentences failing any check were removed. 45.2% of training data survived (237K from 525K).
+
 ## Dataset Formats
 
-### 1. Parquet (recommended for most users)
+### Parquet (recommended)
 
-Auto-detected by HuggingFace. Each row contains:
+Each row contains:
 
 | Column | Type | Description |
 |--------|------|-------------|
@@ -79,9 +85,7 @@ Auto-detected by HuggingFace. Each row contains:
 | `deprels` | list[string] | Dependency relation labels |
 | `ner_tags` | list[string] | BIO-tagged NER labels (18 entity types) |
 
-`prev_text` is populated for 87% of conversation domain sentences (77K of 90K). It contains the preceding utterance from the same conversation, enabling context-aware dialog act classification. For all other domains (business, technical, etc.), `prev_text` is null.
-
-### 2. CoNLL-U (for linguistic tools)
+### CoNLL-U
 
 Standard [CoNLL-U format](https://universaldependencies.org/format.html) with extensions:
 
@@ -102,128 +106,35 @@ Standard [CoNLL-U format](https://universaldependencies.org/format.html) with ex
 10   ?          ?          PUNCT    .      _    5    punct        _    _
 ```
 
-Columns: ID, FORM, LEMMA, UPOS, XPOS, FEATS, HEAD, DEPREL, DEPS, MISC
-
-Metadata comments:
-- `# sent_id` — unique identifier
-- `# text` — original sentence text
-- `# prev_text` — previous utterance (conversation domain only, optional)
-- `# cls` — dialog act classification label
-- `MISC` column — NER tags as `NER=B-ORG`, `NER=I-ORG`, etc.
-
-### 3. Prepared JSON (for direct model training)
-
-Pre-processed splits ready for `torch.utils.data.Dataset`. Located in `prepared/deberta-v3-large-nlp-en/`.
-
-**NER files** (`ner_train.json`, `ner_dev.json`, `ner_test.json`):
-```json
-{
-  "words": ["Apple", "reported", "$394B", "revenue"],
-  "text": "Apple reported $394B revenue",
-  "ner_tags": ["B-ORG", "O", "B-MONEY", "O"],
-  "cls_label": "inform",
-  "prev_text": null
-}
-```
-Sources: kniv corpus (6 domains, spaCy + GPT validated) + GMB (CC BY 4.0, 48K human-annotated).
-
-**UD files** (`ud_train.json`, `ud_dev.json`, `ud_test.json`):
-```json
-{
-  "words": ["The", "cat", "sat"],
-  "text": "The cat sat",
-  "pos_tags": ["DET", "NOUN", "VERB"],
-  "dep_labels": ["-1@det@NOUN", "0@root@ROOT", "+1@nsubj@VERB"],
-  "heads": [1, -1, 1],
-  "deprels": ["det", "root", "nsubj"],
-  "cls_label": "inform"
-}
-```
-Source: UD English EWT v2.14 (CC BY-SA 4.0).
-
-### 4. Gold-validated data (recommended for training)
-
-Located in `prepared/deberta-v3-large-nlp-en-gold/`. This is the same data as above but filtered through **Qwen3-8B** running locally to remove annotation errors.
-
-**Validation process:**
-- Every NER entity span checked: "Is '{entity}' correctly tagged as {type}?" — sentences with any incorrect entity removed
-- Every UD sentence checked: 3 random POS tags validated — sentences with incorrect POS removed
-- ~10.5% of NER sentences removed (bad entity types)
-- ~17.4% of UD sentences removed (ambiguous/incorrect POS)
-
-Result: **36K NER + 10.3K UD gold training examples** with higher annotation confidence than the standard prepared data.
-
-```python
-from huggingface_hub import snapshot_download
-
-# Download gold data (recommended)
-snapshot_download(
-    "dragonscale-ai/kniv-corpus-en",
-    repo_type="dataset",
-    allow_patterns="prepared/deberta-v3-large-nlp-en-gold/*",
-    local_dir="data"
-)
-```
-
-**Label vocabulary** (`label_vocabs.json`):
-```json
-{
-  "ner_labels": ["O", "B-PERSON", "I-PERSON", "B-ORG", "I-ORG", ...],
-  "pos_labels": ["ADJ", "ADP", "ADV", "AUX", ...],
-  "dep_labels": ["+1@nsubj@VERB", "-1@det@NOUN", "0@root@ROOT", ...],
-  "cls_labels": ["inform", "correction", "agreement", ...]
-}
-```
+NER tags in MISC column: `NER=B-ORG`, `NER=I-ORG`, etc.
 
 ## Dataset Details
 
 | | |
 |---|---|
 | **Total sentences** | 657,313 |
+| **Gold sentences** | 237,646 |
 | **NER entity types** | 18 (BIO-tagged = 37 labels) |
 | **POS tags** | 17 (Universal Dependencies UPOS) |
 | **Dep labels** | ~1,440 (dep2label composite tags) |
 | **CLS labels** | 9 (dialog act types) |
-| **Annotation** | spaCy `en_core_web_trf` + GPT-5.4 validation corrections |
-| **CLS annotation** | GPT-5.4-nano with conversational context (prev_text) |
+| **Annotation** | spaCy `en_core_web_trf` + GPT-5.4-nano validation |
+| **Gold filtering** | Qwen3-8B via vLLM (NER + CLS + POS) |
 | **License** | CC BY-SA 4.0 |
 
-### Splits
-
-| Split | Sentences |
-|-------|-----------|
-| train | 525,850 |
-| dev | 65,731 |
-| test | 65,732 |
-
 ### Domains
 
-| Domain | Sentences | Sources | License |
-|--------|-----------|---------|---------|
-| **Conversation** | 89,934 | Taskmaster 1/2/3, OASST1, MultiWOZ 2.2, Glaive Function Calling, Discord Dialogues | CC BY-4.0 / Apache 2.0 / MIT |
-| **Business** | 197,891 | SEC EDGAR 10-K, Enron emails, OpenStax textbooks, Odoo docs, Wikipedia, CUAD contracts, OpenAlex abstracts | Public domain / CC BY-4.0 / CC BY-SA 3.0 / ODC-BY |
-| **Technical** | 193,032 | Wikipedia (CS/engineering), Python documentation | CC BY-SA 3.0 / PSF License |
-| **Encyclopedic** | 66,490 | Wikipedia (science, history, geography, arts) | CC BY-SA 3.0 |
-| **News** | 58,026 | Wikinews, Wikipedia (journalism, politics) | CC BY 2.5 / CC BY-SA 3.0 |
-| **Narrative** | 51,940 | Project Gutenberg (classic fiction) | Public domain |
-
-### Annotation Pipeline
-
-```
-collect → preprocess → annotate (spaCy) → validate (GPT) → classify (GPT) → export
-```
-
-1. **Collect** — raw text from open sources (APIs, HuggingFace datasets, git repos)
-2. **Preprocess** — sentence splitting (non-conversation) or utterance extraction (conversation), cleaning, quality filtering, deduplication, prev_text linking
-3. **Annotate** — spaCy `en_core_web_trf` produces POS, NER, dependency parse per sentence (single pass, both CoNLL-U and JSONL)
-4. **Validate** — GPT-5.4-nano checks POS/NER/dep annotations in batches of 5, returns corrections. Invalid POS corrections rejected (only valid UPOS accepted). Orphan BIO I-tags repaired.
-5. **Classify** — GPT-5.4-nano assigns dialog act labels in batches of 20. Conversation domain sentences include `prev_text` for context-dependent labels (correction, agreement, filler).
-6. **Export** — corrections applied, CLS labels embedded, BIO tags validated, split into train/dev/test, output as CoNLL-U + Parquet
+| Domain | Full | Gold | Sources | License |
+|--------|------|------|---------|---------|
+| **Business** | 197,891 | 73,525 | SEC EDGAR 10-K, Enron emails, OpenStax, Odoo docs, Wikipedia, CUAD contracts, OpenAlex | Public domain / CC BY-4.0 / CC BY-SA 3.0 / ODC-BY |
+| **Technical** | 193,032 | 56,438 | Wikipedia (CS/engineering), Python docs | CC BY-SA 3.0 / PSF License |
+| **Conversation** | 89,934 | 29,203 | Taskmaster 1/2/3, OASST1, MultiWOZ 2.2, Glaive, Discord Dialogues | CC BY-4.0 / Apache 2.0 / MIT |
+| **Encyclopedic** | 66,490 | 29,088 | Wikipedia (science, history, geography, arts) | CC BY-SA 3.0 |
+| **News** | 58,026 | 27,541 | Wikinews, Wikipedia (journalism, politics) | CC BY 2.5 / CC BY-SA 3.0 |
+| **Narrative** | 51,940 | 21,851 | Project Gutenberg (classic fiction) | Public domain |
 
 ### NER Entity Types (18)
 
-BIO-tagged in the `ner_tags` column (Parquet) or MISC field (CoNLL-U):
-
 | Type | Description | Example |
 |------|-------------|---------|
 | PERSON | People | *Caroline*, *Dr. Smith* |
@@ -245,126 +156,88 @@ BIO-tagged in the `ner_tags` column (Parquet) or MISC field (CoNLL-U):
 | LAW | Laws | *the First Amendment* |
 | LANGUAGE | Languages | *English*, *Mandarin* |
 
-NER sources: kniv corpus (spaCy + GPT validated, 6 domains) + GMB Groningen Meaning Bank (CC BY 4.0, 48K human-corrected sentences, 8 entity types mapped to the 18-type scheme).
-
 ### CLS Dialog Act Labels (9)
 
-| Label | Description | Uniko action | Prevalence |
-|-------|-------------|-------------|-----------|
-| `inform` | States a fact, opinion, or observation | Extract observation | ~70% |
-| `question` | Asks for information | Record knowledge gap | ~10% |
-| `request` | Asks someone to do something | Create action node | ~7% |
-| `plan_commit` | Commits to future action, offers, suggests | Link to Goal/Task | ~5% |
-| `correction` | Corrects or contradicts prior statement | Flag knowledge for update | ~3% |
-| `social` | Greeting, goodbye, thanks, apology | Skip extraction | ~3% |
-| `filler` | Turn management, stalling, fragments | Skip entirely | ~3% |
-| `agreement` | Agrees with or confirms something | Reinforce observation | ~2% |
-| `feedback` | Acknowledges without adding information | Skip extraction | ~1% |
-
-Context-dependent labels (correction, agreement, feedback) are classified with the previous utterance (`prev_text`) as context for conversation domain sentences.
+| Label | Description | Uniko action | Gold count |
+|-------|-------------|-------------|------------|
+| `inform` | States a fact or observation | Extract observation | 173,401 (73.0%) |
+| `question` | Asks for information | Record knowledge gap | 14,979 (6.3%) |
+| `plan_commit` | Commits to action, offers, suggests | Link to Goal/Task | 12,518 (5.3%) |
+| `request` | Asks someone to do something | Create action node | 10,679 (4.5%) |
+| `filler` | Fragments, headers, formatting | Skip entirely | 10,538 (4.4%) |
+| `correction` | Corrects or contradicts prior statement | Flag for update | 7,587 (3.2%) |
+| `social` | Greeting, goodbye, thanks, apology | Skip extraction | 3,111 (1.3%) |
+| `agreement` | Agrees with or confirms something | Reinforce observation | 2,956 (1.2%) |
+| `feedback` | Acknowledges without adding info | Skip extraction | 1,877 (0.8%) |
 
 ### dep2label Encoding
 
-Dependency parsing is reformulated as token classification using the [rel-pos encoding](https://aclanthology.org/N19-1077/) (Strzyz et al., 2019). Each token receives a composite label:
+Dependency parsing reformulated as token classification using [rel-pos encoding](https://aclanthology.org/N19-1077/) (Strzyz et al., 2019):
 
 ```
-+1@nsubj@VERB
- │   │     │
- │   │     └── POS tag of the head token
- │   └──────── dependency relation
- └──────────── signed offset (1st VERB to the right)
++1@nsubj@VERB    →  "1st VERB to the right, relation=nsubj"
 ```
 
-Decoding back to a dependency tree is O(n) per sentence.
-
 ## Usage
 
-### Via HuggingFace Datasets (recommended)
+### Load with HuggingFace Datasets
 
 ```python
 from datasets import load_dataset
 
+# Full corpus
 ds = load_dataset("dragonscale-ai/kniv-corpus-en")
 
-# Each row: sent_id, text, prev_text, cls, tokens, pos_tags, heads, deprels, ner_tags
-example = ds["train"][0]
-print(f"Text: {example['text']}")
-print(f"CLS:  {example['cls']}")
-print(f"Prev: {example['prev_text']}")  # None for non-conversation
-for tok, pos, ner in zip(example['tokens'], example['pos_tags'], example['ner_tags']):
-    print(f"  {tok:20s} POS={pos:6s} NER={ner}")
-```
-
-### Training-ready JSON (for model training)
-
-```python
+# Gold corpus (recommended for training)
 from huggingface_hub import snapshot_download
-
-# Download prepared training data
 snapshot_download(
     "dragonscale-ai/kniv-corpus-en",
     repo_type="dataset",
-    allow_patterns="prepared/*",
-    local_dir="data"
+    allow_patterns="corpus/gold/*",
+    local_dir="kniv-data",
 )
 
-import json
-with open("data/prepared/deberta-v3-large-nlp-en/ner_train.json") as f:
-    ner_train = json.load(f)
-print(f"NER training examples: {len(ner_train)}")
-
-with open("data/prepared/deberta-v3-large-nlp-en/label_vocabs.json") as f:
-    vocabs = json.load(f)
-print(f"NER labels: {len(vocabs['ner_labels'])}")
-print(f"POS labels: {len(vocabs['pos_labels'])}")
-print(f"Dep labels: {len(vocabs['dep_labels'])}")
-print(f"CLS labels: {vocabs['cls_labels']}")
-```
-
-### Via CoNLL-U files
-
-```python
-import conllu
-from huggingface_hub import hf_hub_download
-
-path = hf_hub_download("dragonscale-ai/kniv-corpus-en", "train.conllu", repo_type="dataset")
-
-with open(path) as f:
-    sentences = conllu.parse(f.read())
-
-for sent in sentences[:3]:
-    print(f"Text: {sent.metadata['text']}")
-    print(f"CLS:  {sent.metadata.get('cls', 'N/A')}")
-    if 'prev_text' in sent.metadata:
-        print(f"Prev: {sent.metadata['prev_text'][:80]}")
-    for token in sent:
-        ner = token.get("misc", {})
-        ner_tag = ner.get("NER", "O") if isinstance(ner, dict) else "O"
-        print(f"  {token['form']:20s} POS={token['upos']:6s} DEP={token['deprel']:12s} NER={ner_tag}")
+import pandas as pd
+gold = pd.read_parquet("kniv-data/corpus/gold/train.parquet")
+print(f"Gold training: {len(gold)} sentences")
 ```
 
-### Training a multi-task model
+### Train a multi-task model
 
 ```bash
 git clone https://github.com/rustic-ai/kniv-nlp-models
 cd kniv-nlp-models
-uv pip install -e .
 
-# Prepare data (downloads UD EWT + loads corpus NER + GMB)
-uv run python models/deberta-v3-large-nlp-en/prepare_data.py
+# Prepare training data from gold corpus
+python models/deberta-v3-large-nlp-en/prepare_data.py --gold corpus/output/final-gold
 
 # Train teacher (DeBERTa-v3-large, 304M params)
-uv run python models/deberta-v3-large-nlp-en/train.py
+python models/deberta-v3-large-nlp-en/train.py
+
+# Quick test (1 epoch, 100 samples)
+python models/deberta-v3-large-nlp-en/train.py --quick-test
+```
+
+## Annotation Pipeline
 
-# Quick test (1 epoch, 100 samples — verify setup)
-uv run python models/deberta-v3-large-nlp-en/train.py --quick-test
 ```
+collect → preprocess → annotate (spaCy) → validate (GPT) → classify (GPT) → export → gold filter (Qwen3-8B)
+```
+
+1. **Collect** — raw text from open sources (APIs, HuggingFace datasets, git repos)
+2. **Preprocess** — sentence splitting, cleaning, dedup, prev_text linking for conversations
+3. **Annotate** — spaCy `en_core_web_trf` for POS, NER, dependency parse
+4. **Validate** — GPT-5.4-nano checks annotations in batches, returns corrections
+5. **Classify** — GPT-5.4-nano assigns dialog act labels (with prev_text context for conversations)
+6. **Export** — corrections applied, BIO tags repaired, split into train/dev/test
+7. **Gold filter** — Qwen3-8B validates every NER entity, CLS label, and POS tag independently. 64 concurrent requests via vLLM on A100. Sentences failing any check removed.
 
 ## Quality Notes
 
-- **POS/Dep**: Gold-standard from UD English EWT v2.14. GPT validation corrections applied but only valid UPOS tags accepted (invalid GPT suggestions rejected).
-- **NER**: spaCy `en_core_web_trf` annotations + GPT validation corrections + GMB human-corrected data as supplementary source. Orphan I-tags (I- without B-) automatically repaired to B-.
-- **CLS**: GPT-5.4-nano classification. Context-free labels (inform, question, request) are highly reliable. Context-dependent labels (correction, agreement, filler) are improved by prev_text but remain noisier (~60-75% accurate by manual spot-check).
+- **POS/Dep**: UD English EWT v2.14 data used for POS+Dep training head (expert-annotated, not gold-filtered — already gold standard)
+- **NER**: spaCy annotations validated by GPT-5.4-nano, then gold-filtered by Qwen3-8B. ~9.3% of entities rejected.
+- **CLS**: GPT-5.4-nano classification validated by Qwen3-8B. ~4.0% rejected. Context-dependent labels (correction, agreement) remain noisier than context-free ones.
+- **POS on corpus**: spaCy POS validated by Qwen3-8B. ~41% rejected (aggressive filtering — checks 8 tokens per sentence, so even one disagreement drops the sentence).
 
 ## Source
 
@@ -385,4 +258,4 @@ uv run python models/deberta-v3-large-nlp-en/train.py --quick-test
 
 ## License
 
-CC BY-SA 4.0. Individual domain source licenses are listed in the domains table above and in the [source repository](https://github.com/rustic-ai/kniv-nlp-models/tree/main/corpus/domains).
+CC BY-SA 4.0. Individual domain source licenses are listed in the domains table above.