HachimiMT-60-QT

HachimiMT-60-QT: Chinese→Vietnamese translation with a perfectly consistent "convert" voice

🇻🇳 Đọc bản tiếng Việt

The "classical register" edition of HachimiMT-60: it translates Chinese web novels into Vietnamese using the truyện convert pronoun register — ta, ngươi, hắn, nàng, tỷ tỷ, ca ca… (the archaic Sino-Vietnamese address style that Vietnamese web-novel readers know by heart) — and keeps that voice perfectly stable from the first line of a chapter to the last.

What's new in v1.2 (2026-10-23)

Same architecture, same tokenizer, same files — drop-in. v1.2 continues from v1.1 with 2,000 more training steps on 3,444 new sentences (each seen three times), together with the v1.1 phrase sentences and the same 30k replay sentences from the original data:

• 3,143 sentences for 861 term "cells": web slang (打call, 破防, 凡尔赛, 柠檬精 …), chengyu and idioms (卸磨杀驴, 事半功倍, 八字没一撇, 唇亡齿寒 …), words with several senses, kinship terms (姨父, 妯娌 …) and grammar words. Teacher translations were checked by LLM judges, and a QC sample was scored by two independent reviewer agents (3.6 % bad).
• 301 sentences for a subject-role error: when the speaker names themself with a title (本座, 朕, 老夫, 在下, 本王 …) and then says 不要 or 且, v1.1 often read the sentence as an order to the listener ("Trẫm hãy nghe…", "Bổn vương đừng xem").
• 79 replay sentences with known term errors in their reference translation were corrected.

Real output, CTranslate2 INT8, beam 1 (the settings shipped below), one sentence per call:

Measured, v1.1 → v1.2, identical decoding settings:

• Targeted terms: 323 held-out sentences from novels that were never used for training, in the cells above, scored by two LLM judges (a third judge resolves disagreements): error rate 31.8 % → 25.5 % (31 sentences fixed, 11 broken).
• Title + 不要 / 且: 132 probe sentences, wrong reading 72.7 % → 14.4 %.
• Everything else: other cells of the same held-out set 7.0 % → 7.6 % (within noise); 1,368-sentence clean suite BLEU 52.9 → 53.0; repetition rate 0.51 % → 0.58 %; voice consistency still 0 flips in every pronoun class on the 7-chapter test.
• Names: on 13,561 real sentences with 4,518 names from held-out novels, name reading did not change (+0.06 percentage points). On the 230-sentence set that collects fragile rare names on purpose, 224 → 220 (for example 张拙 "Trương Chuyết" can now come out as "Trương Phán").

Still weak in v1.2, so you know: 咱们不要吃 ("let's not eat") now comes out as "Chúng ta không cần ăn" (v1.1: "Chúng ta đừng ăn"); after 师父 / 为师 + 不要 the model says "không cần", which we accept as convert style; 放鸽子 is no longer literal but can still be wrong ("hối hận cho kẻ khác"); rare names can still shift a syllable (苏淡月 → "Tô Nhàn Nguyệt"). Two items from the v1.1 list are fixed in our tests: 敲竹杠 → "vòi vĩnh", and 莫不 after a list of places → "không ai không".

Need earlier weights? Pin a tag: revision="v1.1" or revision="v1.0" (works with from_pretrained and snapshot_download).

What's new in v1.1 (2026-09-12)

Same model, same tokenizer, same files — drop-in. v1.1 was continue-trained for 2,000 steps on 5,380 hand-mined sentences covering "long-tail" phrases the previous version got wrong: verb-object idioms (背黑锅, 敲竹杠, 碰钉子, 钻牛角尖, 狮子大开口, 翻旧账 …), the four-character 宁缺毋滥 (v1.0 turned it into a fake name "Ninh Khuyết"), the literary causative 教 ("khiến", not "dạy"), 莫不 / 讨X / 凛然-type words — each with contrast sentences so the ordinary senses (教 = to teach, names containing 莫/宁) stay intact — plus 30k replay sentences from the original data so nothing else moves.

Real output, CTranslate2 INT8, beam 1 (the settings shipped below):

Measured, v1.0 → v1.1, identical decoding settings:

• Blind review by two independent reviewers, 192 new sentences built around these phrases: mean error severity 0.92 → 0.59; v1.1 better on 46 sentences, worse on 14. On the 175-sentence pack from the first round: 1.01 → 0.50 (53 better / 10 worse).
• Voice consistency (the reason this model exists): still 0 flips in every pronoun class on the 7-chapter test; coverage of the you class 0.92 → 1.00.
• General quality, 1,392-sentence clean suite: BLEU 51.4 → 52.4; repetition rate 0.79 % unchanged; latency unchanged. Rare-name guard (49 hard sentences): BLEU 56.4 → 57.9, name-anchor pass rate unchanged. Name-calibration gate 223/230 unchanged.

Still weak, so you know: 敲竹杠 in some contexts is still literal ("gõ thanh tre"), 放鸽子 is still literal, 莫不 right after a list of places can come out as a name ("Mạc Bất"), and rare names can still shift a syllable or lose capitalisation ("Tô đạm Nguyệt").

Need the previous weights? Pin the tag: revision="v1.0" (works with from_pretrained and snapshot_download).

What's different from the base model?

Ordinary MT models (including base HachimiMT-60) have a habit that drives novel readers crazy: they switch address style mid-chapter. The same character is "ngươi" (archaic you) in one line and suddenly "cậu" (modern you) in the next; the narrator says "ta" (I, archaic) in one paragraph and "tôi" (I, modern) in another.

This QT edition was trained on data whose entire pronoun system was normalized, so the voice never flips. Real output, same chapter, same interview scene (base HachimiMT-60 vs this model):

Measured on 7 chapters across 7 genres (538 lines): in the you-register class the base model flips voice at ~24% of line transitions; this model: 0%. The I, we and third-person classes also drop to zero.

Stress tests

Four pronouns in one sentence — nothing gets mixed up:

他说你不懂她的心思，我们都别插手了。 → Hắn nói ngươi không hiểu tâm tư của nàng, chúng ta đều đừng nhúng tay vào nữa. (he→hắn, you→ngươi, she→nàng, we→chúng ta)

A long multi-clause paragraph — where the base model mixes two voices inside a single paragraph:

张羽同学，我们了解到你的家庭情况恐怕不足以负担这边的学费。不过我们为贫困生提供了优惠的贷款服务，只要你愿意抵押一些不重要的器官就行……

When to use it

• ✅ Xianxia, xuanhuan, historical, danmei — or any genre where you want the convert voice all the way through.
• ⚠️ Modern-setting stories also come out in the archaic voice ("Sao ngươi chưa nghỉ ngơi?") — that is by design. If you want context-appropriate natural register instead, use the base model.

Deliberate trade-offs (documented so nothing surprises you):

• 我们/咱们 both become "chúng ta" (the inclusive/exclusive we distinction is dropped).
• Modern intimate address (anh/em between lovers) mostly becomes ta/ngươi.

Known limitations (real example, not hidden)

Rare proper nouns and coined terms can come out with a shifted syllable, a lower-case syllable or a literal translation — a limitation of every model in this size class, not specific to the QT edition. Examples from v1.2 (CTranslate2 INT8, beam 1): the spirit-beast name 【异瞳金丝猴】 (reference: Dị Đồng Kim Ty Hầu) comes out as "Kim Ti Hầu"; the character 苏淡月 (Tô Đạm Nguyệt) came out as "Tô Nhàn Nguyệt" (v1.1: "Tô đạm Nguyệt"); the Western name 泰尔齐奇 came out half-transliterated ("Til Tề Kỳ"). If your novel is full of first-time rare names, spot-check them.

Quick start (Python)

from transformers import AutoTokenizer, MarianMTModel
import torch

repo = "ngocdang83/HachimiMT-60-QT"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = MarianMTModel.from_pretrained(repo).eval()

src = "你成绩没有问题，但想要进嵩阳高中，光靠校内考试成绩是远远不够的。"
inp = tokenizer(src, return_tensors="pt", truncation=True, max_length=256)
with torch.inference_mode():
    out = model.generate(**inp, max_new_tokens=300, num_beams=4, early_stopping=True)
print(tokenizer.decode(out[0], skip_special_tokens=True))
# "Thành tích của ngươi không có vấn đề gì, nhưng muốn vào trường trung học
#  Tung Dương, chỉ dựa vào điểm thi trong trường thì còn xa mới đủ."

💡 Decoding tip: do not set no_repeat_ngram_size with this model — it forces the decoder to avoid repeating n-grams and tends to mangle proper names (Lý Giáng Thiên → Lý Giáng Dương). The defaults above are what we ship with.

Fast CPU inference (CTranslate2)

The repo ships an INT8 export under ct2-int8_float32/ — several times faster on machines without a GPU:

import ctranslate2
from pathlib import Path
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer

repo = "ngocdang83/HachimiMT-60-QT"
path = Path(snapshot_download(repo, allow_patterns=[
    "config.json", "source.spm", "target.spm", "vocab.json",
    "tokenizer_config.json", "ct2-int8_float32/*",
]))
tokenizer = AutoTokenizer.from_pretrained(path)
translator = ctranslate2.Translator(str(path / "ct2-int8_float32"),
                                    device="cpu", compute_type="int8_float32")

src = "你放心吧，人家是大公司，不会骗人的。"
toks = tokenizer.convert_ids_to_tokens(tokenizer(src).input_ids)
res = translator.translate_batch([toks], beam_size=1, max_decoding_length=256)
print(tokenizer.decode(tokenizer.convert_tokens_to_ids(res[0].hypotheses[0]),
                       skip_special_tokens=True))
# "Ngươi yên tâm đi, người ta là công ty lớn, sẽ không lừa người đâu."

The Hachimi family

Technical details (56.96M params, Marian encoder-decoder, runs fine on CPU or GPU) are in config.json. License CC-BY-4.0 — use freely with attribution.