Add partial multilingual HOSIA expansion

Adds a clearly labeled AI-reviewed draft for two of ten intents (25 aligned scenarios per intent across TR, EN, DE, RU). Not human reviewed or training approved.

PROGRESS.md

@@ -12,16 +12,22 @@ Updated: 2026-09-24
 - Confirmed 100 unique message strings, complete output fields, no email/phone matches in the basic screen, and `reviewed: false` / `allowed_for_training: false` on every record.
 - Drafted the Hugging Face Dataset Card with status, provenance, license restrictions, limitations, and review requirements.
 - Uploaded the Turkish pilot draft, Dataset Card, and progress note to the public `badblli/HOSIA-Intent` dataset repository in commit `4d80e915f79d39725f3a342bf6d89a4c9d1bf04b`; verified all three files through the Hugging Face API.
+- Expanded and AI-reviewed two intents (`extra_towel_request`, `air_conditioning_issue`) into 25 aligned scenarios per intent across TR/EN/DE/RU: 200 records, with normalized entity annotations and scenario-group splits. Translation and intent checks passed; records remain `reviewed: false` and `allowed_for_training: false`.
+- Added resumable multilingual generation, hard-negative generation, and TF-IDF baseline scripts. The current multilingual checkpoint contains only the first 2 of 10 intents.
 
 ## Platform limitation
 
 The EVREN dataset page currently disables the **Metin** modality for this account. Its dataset guide describes image-oriented upload and annotation formats. No EVREN dataset was created because the current text pilot cannot be represented safely in that workflow; the EVREN generation itself is complete and remains in the local audit files.
 
-Hugging Face upload is staged locally, but the `hf` CLI is not authenticated on this machine (`hf auth whoami` reports not logged in, and no cached token or Hugging Face environment token is present). No Hugging Face MCP/upload tool is available in this session. The browser page is open but does not provide an authenticated upload session, and browser file selection was blocked by the computer-use bridge. No remote files have been committed yet.
+Hugging Face CLI is authenticated as `badblli`; the existing public repository contains the pilot, Dataset Card, and progress note. An explicitly named partial file is being uploaded for the two completed multilingual intents. No Hugging Face MCP tool is available in this session.
 
 ## Current artifacts
 
 - `data/generated/evren-v0.1/hosia_v0.1_tr_draft.jsonl` — 100-record Turkish draft; ignored by Git.
+- `data/generated/evren-v0.1/hosia_v0.2_multilingual_checkpoint.jsonl` — 200 AI-reviewed records for the first two intents; ignored by Git; resume generation for the other eight intents.
+- `scripts/build_multilingual_pilot.py` — resumable multilingual generation, entity extraction, validation, and group splitting.
+- `scripts/generate_hard_negatives.py` — separate contrast-set generation pipeline.
+- `scripts/evaluate_tfidf_baseline.py` — development-only word/character TF-IDF baseline.
 - `data/generated/evren-v0.1/candidates.jsonl` — candidate-level model review audit; ignored by Git.
 - `data/generated/evren-v0.1/model_calls.jsonl` — raw generation and validation call audit; ignored by Git.
 - `huggingface/dataset-card.md` — draft card intended for `badblli/HOSIA-Intent`.
@@ -29,6 +35,9 @@ Hugging Face upload is staged locally, but the `hf` CLI is not authenticated on
 ## Outstanding before a training-ready release
 
 - Human review of all examples and correction of any model-label mistakes.
+- Complete multilingual expansion for the remaining eight intents.
+- Generate and validate the separate hard-negative evaluation set.
+- Run the TF-IDF baseline after the full multilingual split exists.
 - Entity annotation; current `entities` fields are empty.
 - Full privacy review beyond email and phone patterns.
 - Hard negatives and English, German, and Russian examples.

README.md

@@ -1,6 +1,9 @@
 ---
 language:
 - tr
+- en
+- de
+- ru
 license: other
 task_categories:
 - text-classification
@@ -9,16 +12,17 @@ size_categories:
 - n<1K
 ---
 
-# HOSIA-Intent — Turkish synthetic pilot (draft)
+# HOSIA-Intent — synthetic hospitality intent data (draft)
 
-This repository currently contains a **100-record Turkish synthetic pilot** for the HOSIA multilingual hospitality intent project. It is a working draft, not a reviewed or training-ready release.
+This repository contains a 100-record Turkish pilot and a clearly labeled partial multilingual expansion for the HOSIA hospitality intent project. It is a working draft, not a reviewed or training-ready release.
 
 ## Status
 
-- 10 pilot intents, 10 records per intent.
-- Candidate labels were checked by a separate EVREN model; this is not human annotation.
-- Records are marked `reviewed: false` and `allowed_for_training: false`.
-- `entities` are empty; scenario groups and evaluation splits are unassigned.
+- `hosia_tr_pilot_draft.jsonl`: 100 Turkish records, 10 pilot intents, 10 per intent.
+- `hosia_v0.2_first_two_intents_partial.jsonl`: 200 AI-reviewed records for two intents, 25 aligned scenarios per intent in TR/EN/DE/RU.
+- EVREN checked candidate labels, extracted entities, and checked translations in the partial expansion; this is not human annotation.
+- Every record is marked `reviewed: false` and `allowed_for_training: false`.
+- The original pilot has empty `entities` and unassigned splits. The partial expansion has normalized entities and scenario-group splits.
 - A basic email and Turkish phone-number scan found no matches. This does not replace full privacy review.
 
 ## Dataset fields
@@ -29,7 +33,7 @@ This repository currently contains a **100-record Turkish synthetic pilot** for
 
 The planned research concerns multilingual hospitality intent classification, entity extraction, department routing, action selection, and information dependency. The current draft should be used only for internal review and iteration. Do not use it for training, evaluation claims, operational decisions, or downstream redistribution.
 
-The examples were generated synthetically through the EVREN LLM API and checked by a separate model. Model review can miss incorrect labels, awkward language, personal data patterns, semantic duplicates, and cultural or language bias. Human review, entity annotation, scenario grouping, leakage-safe splits, and full privacy review remain outstanding. English, German, and Russian records are not included in this pilot despite being planned project languages.
+The examples were generated synthetically through the EVREN LLM API. Model review can miss incorrect labels, awkward language, personal data patterns, semantic duplicates, and cultural or language bias. The multilingual expansion currently covers only two of ten intents. Human review of all records, completion of the other eight intents, hard-negative evaluation, and full privacy and rights review remain outstanding.
 
 ## License and provenance
 
@@ -37,8 +41,7 @@ All rights reserved. No permission is granted to copy, modify, redistribute, tra
 
 ## Next steps
 
-1. Human review and correct each intent label.
-2. Annotate explicit entities and check action and department consistency.
-3. Add hard negatives and cross-lingual examples.
-4. Assign scenario groups and leakage-safe train/validation/test splits.
-5. Complete privacy and rights review before any training-ready release.
+1. Complete the multilingual expansion for all ten intents.
+2. Generate and validate a separate hard-negative evaluation set.
+3. Review model labels/entities and scenario-group splits; complete privacy and rights review.
+4. Run a development baseline, then only prepare a training-ready release after approval.