quannguyen204/diasynth-vi-elderly-orpo-v2

DiaSynth Vietnamese Elderly Care — ORPO

Preference pairs for ORPO finetuning, focused on Vietnamese elderly-care safety hazards. Each record is a single-turn (prompt, chosen, rejected) tuple where chosen is a safe and persona-consistent response and rejected is an unsafe / off-persona variant for the same hazard prompt.

Splits

Schema

{
    "pair_id": "orpo::persona::<hazard>::<seed>::<n>",
    "prompt":   [{"role": "user", "content": "..."}],
    "chosen":   "...",
    "rejected": "...",
    "hazard_category": "<vi_label>",
    "provenance": {}
}

Preprocessing

• —NFC Unicode normalization
• —Token filter: prompt + max(chosen, rejected) ≤ 2048
• —Exact + TF-IDF cosine near-dedup on chosen (threshold = 0.97)
• —Stratified split by hazard_category, group-disjoint by pair_id, ratios = [0.9, 0.09, 0.01] with seed 42

Persona spec

Identical to the SFT dataset (assistant xưng "con").

Intended use

ORPO finetuning (HF TRL ORPOTrainer) with β = 0.1 after SFT, to align the model on safety + persona under elderly-care hazards (phone scams, medication mistakes, etc.).

Limitations

• —Synthetic preference data — chosen/rejected reflect the generator LLM's notion of safe vs unsafe, not a human-labeled gold standard.
• —18 pairs were dropped by the verifier for content drift.
• —Hazard coverage is biased toward the elderly domain; do not use for general-purpose preference alignment.