Native Vietnamese linguist · SFT · RLHF · transcription · evaluation

Your model is only as good as its

I'm Dao Huy (Lucas), a native Vietnamese linguist. I build and grade the data that teaches your model real Vietnamese, SFT pairs, RLHF and preference annotation, prompts, transcription and evaluation, with the judgement crowd labellers miss, and a written rationale on every label.

500+
hours annotated
Scale AI · Mindrift
trusted by
4.99★
Upwork, 81 reviews
Who
Native Vietnamese linguist, Da Nang, 7+ years
Services
SFT · RLHF / preference · prompts & red-team · transcription / ASR · evaluation · linguistic QA
Proof
500+ hours · Scale AI (200+) · Mindrift / Turing (300+) · Upwork Top-Rated Plus
Pairs
English · Chinese · French → Vietnamese
Formats
JSONL · CSV · XLIFF · CoNLL · ELAN / TextGrid · your schema
ngang · native judgement

Fluent Vietnamese is easy. Knowing when it's wrong is the job.

A tone mark is the difference between a friend and a grave. Tap a syllable to walk its meanings, the only thing changing is the dấu.

Why Vietnamese is hard for AI

Six places a model trained on scraped or crowd-labelled Vietnamese quietly goes wrong, and where native judgement earns its keep.

Tonema · má · mà · mả · mã · mạ

Six tones, six different words on one syllable. A model that flattens tone ships confident nonsense.

Diacriticscà phê → ca phe

Strip the dấu and the word changes or dies. Scraped Vietnamese is often tone-stripped, which quietly poisons training data.

Register & honorificsem · anh · chị · dạ

Pronouns encode rank and age. Flatten "em/anh" to "I/you" and the hierarchy a native always hears is gone.

DialectsBắc · Trung · Nam

Northern, Central and Southern differ in lexicon and tone. "Correct" depends on the target audience.

Classifierscon · cái · chiếc

Vietnamese needs the right classifier per noun. Models guess; a native catches it instantly.

Code-switching"chốt deadline nhé"

Real Vietnamese mixes English tech terms. Knowing when to keep, gloss or translate is judgement, not a rule.

sắc · how I grade

Every rejection comes with a reason.

The same loop I run for Scale AI and Mindrift: read the prompt, compare outputs, choose, and write down why, so the preference data is auditable, not a vibe.

huyền · register

One message, four registers.

"I need two days off." Vietnamese encodes the relationship in every pronoun and particle. Switch the register and watch the same intent change shape.

hỏi · the six tones

Six tones on one syllable.

"ma" carries six different words depending on its tone contour. Click a tone to draw its pitch and hear a stylised version.

ngã · spot the native

Which one would a native ship?

Three rounds. Pick the output you'd accept into a Vietnamese dataset. The rationale and verdict reveal only after you choose.

nặng · field notebook

Seven years of reading Vietnamese closely.

Hover the marked terms for the working note behind them. This judgement is what I bring to a label spec.

Hover a marked term →

Native expert vs crowd vs synthetic

What separates data a model can trust from data that teaches it confident mistakes.

Quality signalNative expert (me)CrowdsourcedSynthetic / scraped
Register & honorificsControlledOften wrongFlattened
False friendsCaughtMissedAmplified
FactualityVerifiedVariesHallucinated
Diacritic integrityIntactVariesOften stripped
Rationale per labelEvery itemNoneNone
Consistency at scaleOne standardInter-rater driftUniform but wrong
process · how the data gets made

From spec to graded data.

The same loop whether it is a fifty-item calibration set or a five-hundred-hour programme.

1

Scope & guidelines

We align on the task, the label spec, the schema and an edge-case rubric. I flag the ambiguities before a single label is written.

2

Calibration batch

A small pilot you review, so the standard is locked before scale. Every disagreement becomes a written rule, not a guess repeated a thousand times.

3

Production with rationale

Data authored or graded at volume, each item carrying the reason behind it, so quality stays auditable instead of a black box.

4

QA & delivery

A consistency pass across the whole batch, then delivery in your format with a short error report. Revisions until it is clean.

quote · working together

Send a task spec, get a plan in a day.

No fixed menu. Tell me the task and I scope it to your guidelines.

Tell me the task, the language pairs, the volume and your schema. You get back an approach, a rate and a calibration plan, usually within one business day.

SFT / instructionRLHF / preferencePrompts & red-teamTranscription / ASREvaluation & rubricsLinguistic QA

Pricing: hourly or per item, locked after a short paid calibration batch · NDA before any data · Reply within a business day · USD via Upwork, bank transfer, PayPal, Wise.

faq

Frequently asked.

glossary

The terms, in plain words.

get a quote

Send me a sample. I'll grade it and tell you what your labellers missed.

NDA before any data · reply within a business day · USD via Upwork, bank, PayPal, Wise

Email a sample to grade →