PROGRESS.md

HOSIA Dataset Generation Progress

Updated: 2026-09-25

v0.3 intent expansion

• The ontology already defined 20 hospitality intents; the earlier dataset config and multilingual draft used only the first 10. Added data/config/hosia_v0.2_intents.json to align the dataset config with all 20 ontology intent IDs and clarify close intent boundaries.
• Generated the remaining 10 intents with EVREN and extended the multilingual draft to 20 intents × 25 scenarios × 4 languages: 2,000 records across 500 scenario groups. The group-safe split has 1,680 train, 160 validation, and 160 test records.
• Turkish candidates were generated with gemma-4-31b and qwen3.8-flash-next, translated with Gemma/Qwen, and primarily reviewed with deepseek-v4.1-flash. The audit contains 118 successful calls and 485,500 provider-reported tokens, including translation repairs and a duplicate repair.
• One repeated Russian phrasing was rewritten with EVREN and revalidated. The extension generator now recovers successful candidate batches from its ignored audit log after an interrupted run and records actual generator/translator models.
• Ontology validation passed (40 concepts, 40 relations, 80 lexical mappings); the 2,000-row combined dataset passed the balanced v0.2-intent validator. All records remain reviewed: false and allowed_for_training: false; AI review is not human review.
• The public hard-negative set still covers only the original 10 intents and is not an undisclosed benchmark. The new 20-intent dataset draft needs a matching hard-negative review set before model-quality conclusions are broadened.
• Local output: data/generated/evren-v0.1/hosia_v0.3_multilingual_ai_reviewed.jsonl (ignored by Git). The v0.3 data, 20-intent config, and Dataset Card were published to badblli/HOSIA-Intent in HF commit 3dc057b33b9c1a1ad79fd977d87f10dc1d0c3c53.

Quality harness update

• Added requirements-dev.txt with pytest and wired GitHub Actions to run unit tests plus ontology validation.
• Added requirements-ml.txt for the optional scikit-learn baseline dependency so CI stays lightweight.
• Added tests/test_dataset_validator.py and scripts/validate/validate_dataset.py to validate generated main and hard-negative JSONL drafts locally. Dataset artifacts remain excluded from Git, so CI uses small synthetic fixtures; the full local datasets were also checked directly.
• Latest checks: 8 pytest tests passed; ontology validation passed (40 concepts, 40 relations, 80 lexical mappings); main draft passed (1,000 rows, balanced 10 intents × 4 languages, 840/80/80 group-safe splits); hard-negative draft passed (80 rows).

Hard-negative audit

• Audited all 80 contrast examples with blind, shuffled intent classification from the three models currently listed by EVREN (deepseek-v4-flash, glm-5.3, qwen3.8-flash-next). The models did not see the current label or contrast target.
• The 2-of-3 vote matched the current label on 80/80 examples; all three models agreed with the current label on 74/80. Six rows had one dissenting model. No labels were changed; model agreement is not human ground truth.
• The TF-IDF baseline classified 31/80 contrast examples incorrectly, while the independent EVREN majority supported their current labels. This points to difficult distinctions in the challenge set, not automatic proof of bad labels.
• The challenge set and labels are public in this draft, so it must not be treated as an undisclosed benchmark holdout. A future benchmark needs a new, properly reviewed private/held-out challenge set.
• Full local report and raw model audit are under ignored data/generated/evren-v0.1/hard_negative_blind_audit.* files.

Completed

• Expanded the hospitality ontology from 4 to 20 intents, with department/action relations and TR/EN/DE/RU lexical mappings.
• Generated a 100-record Turkish pilot through EVREN and AI-checked its candidate intent labels.
• Completed a multilingual draft for 10 intents × 25 scenarios × 4 languages: 1,000 records across 250 scenario groups. Each intent-language pair has 25 records.
• Kept scenario groups together across train/validation/test splits: 840 train, 80 validation, 80 test. All records remain reviewed: false and allowed_for_training: false.
• Ran EVREN-assisted label/entity checks on Turkish sources and translation checks/repairs on multilingual records. This is model review, not human annotation.
• Generated a separate 80-record hard-negative challenge set (10 contrast pairs × 2 examples × 4 languages); labels were checked by EVREN. It is marked hard_negative_eval, unreviewed, and training-disabled.
• Ran a development-only word/character TF-IDF + LinearSVC baseline: test accuracy 0.988, macro-F1 0.987; hard-negative foil accuracy 0.613. These synthetic-pilot metrics are not publishable benchmark claims.
• Basic email and phone-pattern screening found no matches. This does not replace full privacy review.
• Configured an EVREN API key locally for selected generation and validation models; the key itself is not stored in this repository. Accepted EVREN LLM Gateway terms v1 after user approval.
• Uploaded the final full draft and reports to the public Hugging Face dataset repository badblli/HOSIA-Intent (commit f9391a7e188367ec63c9eb858d6d0ac24f6b5f43). The superseded pilot and partial JSONL files were removed from the main branch to avoid duplicate ingestion; they remain in commit history.

Platform limitation

The EVREN dataset page currently disables the Metin modality for this account. Its available upload workflow does not safely represent this text dataset, so no EVREN dataset was created. EVREN was used for generation and model-assisted validation.

Current artifacts

• data/generated/evren-v0.1/hosia_v0.2_multilingual_ai_reviewed.jsonl — final 1,000-record multilingual draft; ignored by Git.
• data/generated/evren-v0.1/hosia_v0.2_hard_negatives.jsonl — 80-record hard-negative challenge set; ignored by Git.
• data/generated/evren-v0.1/tfidf_baseline.md and .json — development baseline report and metrics.
• data/generated/evren-v0.1/candidates.jsonl, model_calls.jsonl, multilingual_review_calls.jsonl, hard_negative_calls.jsonl — ignored local generation/review audit trails.
• scripts/build_multilingual_pilot.py — resumable generation, extraction, validation, and group splitting.
• scripts/extend_multilingual_intents.py — resumable expansion from 10 to all 20 ontology intents, multilingual translation/review, duplicate repair, and combined draft assembly.
• data/config/hosia_v0.2_intents.json — 20 aligned canonical intent definitions.
• scripts/generate_hard_negatives.py — contrast-set generation and validation.
• scripts/audit_hard_negatives.py — resumable blind second-opinion audit across available EVREN models.
• docs/hard-negative-audit-summary.md — vote totals, dissenting records, and provisional label boundaries.
• scripts/evaluate_tfidf_baseline.py — development-only baseline.
• scripts/validate/validate_dataset.py, tests/test_dataset_validator.py, requirements-dev.txt — pytest and JSONL validation harness.
• huggingface/dataset-card.md — dataset card source.

Remaining before a training-ready release

• Human review of examples, intent labels, entity annotations, translations, and hard negatives.
• Entity review and completion where needed; empty entities can be valid for intents without slots, but all cases need a schema-aware review.
• Full privacy review beyond basic email/phone patterns, including indirect identifiers and sensitive content.
• Rights and release review. All rights remain reserved; publication grants no reuse permission.
• Revisit EVREN dataset upload if text modality or a text import path becomes available.
• Consider adding more intents and examples after review of the current ontology-aligned 10-intent scope.

Next milestone

Complete human/privacy/rights review, inspect hard-negative failure modes and annotation gaps, then decide whether this draft can be prepared for any training or benchmark use.