laya-vi

A Vietnamese fine-tune of Laya. Give it a state (an email, report, chat log or document) and typed questions, and it returns typed answers with calibrated probabilities in a single forward pass. Like Laya, it never generates text, so there is nothing to parse and nothing to hallucinate.

laya-vi starts from the laya-multilingual checkpoint and is trained on Vietnamese business decisions. On Vietnamese decisions from the domains it was trained on, accuracy rises from 0.45 to 0.89. On domains held out from training it also improves, and its probabilities are much better calibrated than the base checkpoint's.

Installation

pip install laya

laya-vi uses the standard Laya checkpoint format, so the Laya runtime loads it with no extra code.

Quickstart

import laya

agent = laya.load("vanty120/laya-vi")

state = ("Chào shop, tôi đã chuyển khoản 2 lần cho đơn hàng #4521 tháng 3. "
         "Nhờ shop hoàn lại khoản bị trừ trùng trong hôm nay, nếu không tôi sẽ huỷ tài khoản.")
questions = {
    "bo_phan": {"type": "choice", "instructions": "Bộ phận nào nên xử lý yêu cầu này?",
                "criteria": {"thanh_toan": "hoá đơn, thanh toán, hoàn tiền",
                             "ky_thuat": "lỗi, sự cố, hệ thống",
                             "khac": "các vấn đề khác"}},
    "muc_do_khan": {"type": "score", "instructions": "Yêu cầu này khẩn cấp đến mức nào?",
                    "criteria": ["không gấp", "sớm", "rất gấp"]},
    "nguy_co_roi_bo": {"type": "noul", "instructions": "Khách hàng có đe doạ huỷ hoặc rời bỏ dịch vụ không?"},
}

result = agent.predict(state, questions)
print(result["answers"]["bo_phan"]["choice"])        # thanh_toan   (p = 0.96)
print(result["answers"]["muc_do_khan"]["score"])     # 1.71 on a 0–2 scale ("rất gấp": p = 0.75)
print(result["answers"]["nguy_co_roi_bo"]["noul"])   # 0.94  (probability the answer is yes)

The three question types:

Type	Answer
noul	probability that the answer is yes
choice	one of the listed options (3–10), with a probability for each
score	expected position on an ordinal scale (3–5 levels), with a probability for each level

It also works with laya-serve and the other Laya integrations. See the Laya documentation for details.

Results

Test set	n	Accuracy (laya-multilingual → laya-vi)	ECE (laya-multilingual → laya-vi)
Trained domains, unseen cases	13,130	0.452 → 0.890	0.338 → 0.081
Held-out domains (zero-shot)	6,220	0.498 → 0.596	0.290 → 0.091
MASSIVE vi-VN (human labels)	1,000	0.520 → 0.571	0.218 → 0.134

n counts (document, question) pairs. laya-multilingual is the original Laya checkpoint, evaluated zero-shot. ECE is expected calibration error, where lower means the probabilities better match observed accuracy. Brier score, KL, score MAE and other metrics are in eval_results.json.

Median latency is about 24 ms per request on a single GPU.

Training data

The training data is Vietnamese and fully synthetic. No real customer or user data was used. It covers everyday decisions in Vietnamese business settings:

• customer-service emails
• invoice reconciliation
• power-grid incident reports
• occupational safety
• tender and bidding documents
• recruitment
• product reviews
• comment moderation
• project team group chats
• subcontracting agreements

The documents use Vietnamese names, companies, currency and workplace conventions. About a quarter of them also appear without diacritics (không dấu), as Vietnamese is often typed that way. Two of the domains are held out of training and used only to measure zero-shot performance.

Documents are generated from known underlying facts. Each case is labeled by an LLM teacher three times, and the gold label is the average of the three answer distributions. Automatic checks remove cases where the label contradicts the facts the document was built from, near-duplicates and malformed text.

Training procedure

• 1 epoch of soft cross-entropy against the teacher distributions, then 3 epochs of RLCD (Laya's reinforcement-learning objective based on proper scoring rules). English replay makes up 15% of training, to limit forgetting.
• The best epoch is chosen on a separate calibration split. Temperatures are then fitted per question type on that split.
• About 1.8 hours on a single NVIDIA GB10. The full configuration is in train_log.json.

Limitations

• Gains are largest on the domains in the training data. On new domains, expect results closer to the held-out row above, and fine-tune on your own decisions where accuracy matters.
• Labels come from an LLM teacher, so the model can inherit that teacher's mistakes and biases.
• It is not a substitute for human judgement in safety-critical decisions.

License and acknowledgements

Apache 2.0, the same license as the base model. laya-vi is built on Laya by Convai Innovations (GitHub). Please credit Laya when you use this model.