Doreen AI | Multimodal Data Studio

ID	Domain	Sub-Topic	English Source	Gold Standard	Audio	QA	Status	Action
Loading records...

ID	Domain	Sub-Topic	English Word	Burmese Translation	Audio	Status	Action
Loading records...

✨ Native Gold Standard (Focus Mode)

Linguist Transcreation Guide

General Rule: Transcreation > Translation

Never translate word-for-word. Read the English source, understand the intent using the Literal Meaning box, and rewrite it into a natural, culturally authentic Burmese sentence.

🩺 1. Medicine (Clinical)

Objective: Train the AI for telehealth, patient intake, and medical transcriptions.

Characteristics: Focuses on realistic clinical dialogue (SOAP notes, symptom intake).
Tip: Keep standard medical acronyms (BP, Paracetamol, OD, mg) in English or Latin. Do not invent awkward Burmese translations for standard clinical terms.

🎓 2. Education & Language

Objective: Teach the model to act as a culturally accurate language tutor.

Characteristics: Covers greetings, travel, and classroom questions.
Tip: Strict enforcement of polite honorific particles (ပါ/ရှင်/ခင်ဗျာ). Grammatical etiquette must be pristine.

💳 3. Finance & Banking

Objective: Equip the model for fintech and mobile banking support.

Characteristics: Mobile wallet transfers, KYC, fraud warnings.
Tip: Use everyday Myanmar financial terminology (e.g., ငွေလွှဲခြင်း for transfer, အတိုးနှုန်း for interest rate). Sound professional and secure.

💬 4. General Conversation

Objective: Build baseline empathy and everyday customer support tone.

Characteristics: Polite apologies, de-escalation, expressing gratitude.
Tip: Zero technical jargon. Focus on organic flow and adapt the tone to be highly respectful, treating the user as a valued client.

🔮 5. Prophet AI (Spiritual)

Objective: Generate highly authentic astrological counseling (Baydin).

Characteristics: Covers BaZi, Palmistry, Face Reading, Mahabote, and Tarot.
Tip: Employ the "Law of Duality." Give brutally honest but polite critiques. Avoid purely positive fluff. Use natural Asian idioms instead of robotic translations.

🎙️ 6. Acoustic STT Rules (The One-Second Rule)

Objective: Record clean, perfectly padded audio for the Speech-to-Text engine.

The Rule: Click Record ➔ Wait 1 Second ➔ Speak naturally ➔ Wait 1 Second ➔ Click Stop.
Warning: If you start speaking the instant you click record, your first consonant will be cut off. If you wait 3+ seconds before stopping, the AI will hallucinate timestamps. The system will auto-fail your audio if you violate this pacing!

📦 7. Pipeline Verification Summary & Exporter Reference

Below is the technical specification of each export button available across the studio:

Button	Endpoint Called	Target Format	Description / Engine Target
Refresh	Frontend Event	Live Database Sync	Reloads the vault table and updates category metrics without re-exporting.
SFT Gemini	/api/studio/export_jsonl?format=vertex	Gemini Conversation Text SFT	Exports sentence transcreations as `{"role": "user"}` & `{"role": "model"}` for Google Vertex AI fine-tuning.
SFT Gemma	/api/studio/export_jsonl?format=chatml	Gemma ChatML Text SFT	Exports conversational text with `<\|im_start\|>user/assistant` markers for open-source LLMs.
STT Gemini	/api/studio/lexicon/export?format=stt_vertex	Gemini Multimodal Audio (`contents/fileData`)	Exports audio manifests with `gs://...` URIs and transcription prompts for Gemini Multimodal training.
STT Whisper	/api/studio/lexicon/export?format=stt_whisper	Open-Source (`audio_path/transcript`)	Manifest linking cloud audio files to text for OpenAI Whisper and Hugging Face ASR pipelines.
TTS Piper	/api/studio/lexicon/export?format=tts_piper	Piper LJSpeech (`audio_filepath/phonemes`)	Exports acoustic manifests paired with phonetic pronunciations for training on-device edge TTS models.
Lexicon Gemini	/api/studio/lexicon/export?format=lexicon_vertex	Vocabulary SFT for Gemini	Exports technical glossary definitions and transcreations formatted for Vertex AI dictionary tuning.
Lexicon Gemma	/api/studio/lexicon/export?format=lexicon_chatml	Vocabulary SFT for Gemma	Exports domain-specific terms in ChatML format for Hugging Face open-source token tuning.

Multimodal Exporters & Verification Summary

Multi-Target Exporter Overview

Every approved sentence and audio take in Doreen AI Studio can be exported into multiple formats for downstream model training.

SFT Gemini

Creates a Vertex AI JSONL file mapping English queries to Native Gold Burmese responses using {"role": "user"} and {"role": "model"}.

SFT Gemma

Creates an Open-Weights ChatML JSONL dataset formatted with <|im_start|> tokens for fine-tuning Gemma and LLaMA.

STT Gemini

Generates a Multimodal Speech-to-Text manifest combining both Sentence and Lexicon cloud audio (gs://...) into Vertex fileData structures.

STT Whisper

Produces an open-source manifest with audio_path and transcript keys for OpenAI Whisper training scripts.

TTS Piper

Exports an acoustic manifest with phonetic notes for low-latency, on-device Text-to-Speech synthesis models.

Refresh

Queries PostgreSQL to reload the table and refresh live counts without downloading files.

Verification Summary Table

Button	Endpoint Called	Target Format
SFT Gemini	/api/studio/export_jsonl?format=vertex	Gemini Conversation Text SFT
SFT Gemma	/api/studio/export_jsonl?format=chatml	Gemma ChatML Text SFT
STT Gemini	/api/studio/lexicon/export?format=stt_vertex	Gemini Multimodal Audio `contents/fileData`
STT Whisper	/api/studio/lexicon/export?format=stt_whisper	Open-Source `audio_path/transcript`
TTS Piper	/api/studio/lexicon/export?format=tts_piper	Piper LJSpeech `audio_filepath/phonemes`
Lexicon Gemini	/api/studio/lexicon/export?format=lexicon_vertex	Vocabulary SFT for Gemini
Lexicon Gemma	/api/studio/lexicon/export?format=lexicon_chatml	Vocabulary SFT for Gemma