---
license: cc-by-nc-4.0
language: vi
task_categories:
- text-generation
tags:
- vietnamese
- elderly-care
- synthetic
- preference
- orpo
- safety
pretty_name: DiaSynth Vietnamese Elderly Care — ORPO
size_categories:
- 1K<n<10K
---

# 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

| Split | Pairs | Bytes |
|---|--:|--:|
| train | 14,927 | 28,547,458 |
| val   | 1,494   | 2,867,367 |
| test  | 165  | 314,853 |

## Schema

```python
{
    "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.