---
license: apache-2.0
task_categories:
- text-generation
- question-answering
language:
- vi
tags:
- vietnamese
- sft
- instruction-tuning
- conversational
- multi-turn
- reasoning
- llm-as-a-judge
- unsloth
- curriculum-learning
- synthetic-data
- data-centric-ai
- preview-dataset
pretty_name: "Vietverse-SFT (Preview): A High-Density Native Vietnamese SFT Dataset"
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: train
path: "data/train-*.parquet"
---
# 🇻🇳 Vietverse-SFT v1.0 (Preview Edition)
### *A High-Density, Native Vietnamese Foundation Dataset for Supervised Fine-Tuning*
### *Bộ Dữ Liệu Nền Tảng SFT Tiếng Việt Bản Xứ Chuẩn Mực Cho Huấn Luyện Mô Hình Ngôn Ngữ Lớn*
🇻🇳 [Đọc Bản Tiếng Việt] •
🇬🇧 [Read English Version]
[](#)
[](https://opensource.org/licenses/Apache-2.0)
[](#)
[-F59E0B?style=for-the-badge&logo=star&logoColor=white)](#)
[](#)
[](#)
[**🤗 Hugging Face Dataset**](https://huggingface.co/datasets/TTP01/Vietverse-sft-preview) • [**⚡ Hướng Dẫn Sử Dụng / Quickstart**](#quickstart)
---
| 🌐 Ngôn Ngữ / Language | 📌 Chuyển Hướng / Quick Navigation |
| :--- | :--- |
| **🇻🇳 Bản Tiếng Việt Đầy Đủ** | 👉 [**Xem toàn bộ Báo cáo Kỹ thuật Tiếng Việt**](#vietnamese) |
| **🇬🇧 Full English Version** | 👉 [**Jump to the complete English Technical Dataset Card**](#english) |
---
---
🇻🇳 PHẦN 1: TIẾNG VIỆT (BẢN BÁO CÁO KỸ THUẬT CHI TIẾT)
> 📢 **THÔNG BÁO: BẢN PREVIEW 100 MẪU THỬ NGHIỆM (COMMUNITY DEMO)**
>
> Đây là bản **Preview 100 mẫu chuẩn Gold Tier (≥ 9.0/10.0)** đầu tiên của dự án **Vietverse-SFT**.
> - **Mục đích**: Chia sẻ cho cộng đồng AI / LLM Việt Nam tải về test thử nghiệm thực tế về phong cách hội thoại tiếng Việt tự nhiên, độ sâu kiến thức và cấu trúc metadata 35 trường.
> - ☕ **Bản Full Release**: *Hiện tại chưa có kế hoạch hoặc thời gian phát hành cụ thể cho bản Full (khi nào có thêm kinh phí và thời gian sẽ làm tiếp sau :)))).*
---
## 1. Bối Cảnh & Động Lực Nghiên Cứu (Motivation)
Trong quá trình huấn luyện các mô hình ngôn ngữ lớn (LLMs) tiếng Việt, vấn đề nan giải nhất hiện nay là **sự suy giảm chất lượng do phụ thuộc vào các bộ dữ liệu dịch máy từ tiếng Anh**. Việc dịch thô từ các tập dữ liệu phương Tây (Alpaca, ShareGPT, UltraChat dịch tự động) để lại những hệ quả nghiêm trọng:
1. **Hiện tượng dịch máy gượng gạo (Translation Artifacts)**: Câu hỏi bị rập khuôn bằng các mẫu từ ngữ sách vở như *"Hãy...", "Bạn hãy...", "Vui lòng cho tôi biết..."* — những cấu trúc mà người dùng Việt Nam thực tế không bao giờ sử dụng khi trò chuyện với AI.
2. **Mất liên kết văn hóa & bản xứ (Cultural Disconnect)**: Thiếu vắng các ngữ cảnh đời thực của Việt Nam (thủ tục hành chính, pháp luật, y tế, giáo dục, đời sống văn hóa xã hội).
3. **Loạn xưng hô (Pronoun Confusion)**: AI dịch máy thường lẫn lộn xưng hô, xưng *"tôi/em/chúng tôi"* một cách bất nhất.
**Vietverse-SFT** được xây dựng dựa trên nguyên tắc **Data-Centric AI (Chất lượng hơn số lượng)** — lấy cảm hứng từ các nghiên cứu nền tảng như *LIMA (Zhou et al., 2023)* và *Alpagasus (Chen et al., 2024)*, chứng minh rằng một tập dữ liệu nhỏ nhưng có độ tinh khiết cao, đa dạng persona và được kiểm định chặt chẽ sẽ đem lại hiệu năng hội thoại vượt trội so với các tập dữ liệu rác hàng trăm nghìn mẫu.
---
## 2. Kiến Trúc Pipeline 3-Stage Multi-Agent Synthesis
Toàn bộ dữ liệu của Vietverse-SFT không được sinh ngẫu nhiên theo kiểu zero-shot, mà được vận hành thông qua quy trình **Multi-Agent 3 Giai Đoạn** có kiểm soát nghiêm ngặt:
```
┌────────────────────────────────────────────────────────────────────────────────────────┐
│ VIETVERSE-SFT 3-STAGE MULTI-AGENT PIPELINE │
├──────────────────────────┬──────────────────────────┬──────────────────────────────────┤
│ STAGE 1: USER SIMULATOR │ STAGE 2: DOMAIN EXPERT │ STAGE 3: G-EVAL QUALITY JUDGE │
│ (Mô phỏng Người Dùng) │ (Chuyên gia Trả Lời) │ (Giám Khảo Độc Lập) │
├──────────────────────────┼──────────────────────────┼──────────────────────────────────┤
│ • 19 Persona thực tế │ • Mật độ thông tin cao │ • Chấm 5 tiêu chí (1.0 - 5.0) │
│ • Phương ngữ Bắc/Trung/Nam│ • Văn phong tự nhiên │ • Trỏ đúng bằng chứng trong văn bản│
│ • Evol-Instruct ràng buộc│ • Chống nịnh bợ (Anti) │ • Triệt tiêu Verbosity & Halo bias│
│ • Cấu trúc mở đầu mở │ • Xưng hô chuẩn mực │ • Quality Gate nghiêm ngặt │
└──────────────────────────┴──────────────────────────┴──────────────────────────────────┘
```
### Chi Tiết Từng Giai Đoạn:
1. **Stage 1 — User Simulator (Bộ Mô Phỏng Người Dùng)**:
- Áp dụng phương pháp *Persona Hub (Tencent AI Lab)*: Trước khi sinh câu hỏi, hệ thống ngầm xây dựng chân dung người dùng (vai trò, mức độ am hiểu, trạng thái tâm lý, thiết bị gõ).
- Áp dụng kỹ thuật *Evol-Instruct*: Tăng độ phức tạp bằng cách bổ sung ràng buộc thực tế (ngân sách, lỗi code cụ thể, phiên bản phần mềm, quy định địa phương).
- Đa dạng hóa mở đầu: Triệt tiêu hoàn toàn thói quen lặp mẫu *"Mình đang..."* — hỗ trợ hỏi trực diện, nêu tình huống, kiểm chứng niềm tin có sẵn.
2. **Stage 2 — Domain Expert Assistant (Chuyên Gia Trả Lời)**:
- Trả lời súc tích, giàu tri thức, trực tiếp giải quyết vấn đề.
- **Anti-Sycophancy (Chống nịnh bợ)**: Khi người dùng đưa ra nhận định hoặc giả định sai lầm (*"Nghe nói cứ tăng RAM là hết lag..."*), mô hình sẵn sàng chỉ ra điểm chưa chính xác một cách tôn trọng và khoa học, không đồng tình vô căn cứ.
- Quy tắc xưng hô nhất quán: Duy trì chuẩn mực trò chuyện 1-on-1 tự nhiên (`mình/tôi - bạn`).
3. **Stage 3 — G-Eval Quality Judge (AI Giám Khảo Độc Lập)**:
- Chấm điểm cặp hội thoại theo 5 tiêu chí rõ ràng dựa trên **bằng chứng cụ thể xuất hiện trong văn bản**.
- Triệt tiêu thiên kiến chuộng câu trả lời dài (*Verbosity Bias*) và thiên kiến hào nhoáng bề ngoài (*Halo/Authority Bias*).
---
## 3. Phân Bổ Năng Lực & Dữ Liệu (Taxonomy & Distribution)
### 📊 A. Phân Bổ Thực Tế Của Bản Preview 100 Mẫu Gold (Empirical 100-Sample Stats)
Toàn bộ **100 mẫu trong bản Preview** đều đã được kiểm tra, xác thực và phân bổ đa dạng qua các lĩnh vực thực tế:
```
Education (Giáo dục & Học đường) ████████████████████ 20% (20 mẫu)
Knowledge VN (Tri thức bản địa) ██████████████████ 18% (18 mẫu)
Economy & Finance (Kinh tế) ████████████████ 16% (16 mẫu)
Law VN (Pháp luật Việt Nam) ████████████████ 16% (16 mẫu)
Technology & Code (Công nghệ) ███████████████ 15% (15 mẫu)
General Chat (Đàm thoại tư vấn) ████████ 8% (8 mẫu)
Healthcare (Y tế & Đời sống) ███████ 7% (7 mẫu)
```
| Tiêu Chí Thống Kê | Số Liệu Thực Tế Trong 100 Mẫu Preview |
| :--- | :--- |
| **🥇 Tỷ Lệ Gold Tier (≥ 9.0/10.0)** | **100% (100/100 mẫu đạt chuẩn Gold)** |
| **⭐ Điểm Chất Lượng Trung Bình (Quality Score)** | **`9.63 / 10.0`** (Min: `9.0`, Max: `10.0`) |
| **📏 Độ Dài Trung Bình Mỗi Mẫu (Total Tokens)** | **`928.9 tokens / sample`** |
| **🗺️ Phân Bổ Phương Ngữ (Dialect)** | Bắc (`north`): **49%** • Nam (`south`): **48%** • Trung (`central`): **3%** |
| **🎯 Phân Bổ Độ Khó (Difficulty)** | `expert`: **29%** • `easy`: **29%** • `hard`: **21%** • `medium`: **21%** |
| **🔄 Phân Bổ Lượt Hỏi (Turn Count)** | 1-turn: **90 mẫu** • 2-turn: **4 mẫu** • 3-turn: **3 mẫu** • 4-turn: **3 mẫu** |
---
### 🗺️ B. Phân Bổ Danh Mục Thiết Kế (Target Taxonomy Overview)
```
Reasoning & Logic ██████████████████████ 22%
Core Chat & Advice ████████████████████ 20%
Code & Technical ███████████████ 15%
Multi-Turn Dialogue ████████████ 12%
Tool Use & Agentic ██████████ 10%
Domain Knowledge ██████████ 10%
Safety & Alignment ██████ 6%
Creative Writing █████ 5%
```
| Nhóm Năng Lực (Shard) | Tỷ Lệ | Trọng Tâm Kiến Thức & Kỹ Năng |
| :--- | :---: | :--- |
| 🧠 **Reasoning & Logic** | **22%** | Toán học, suy luận logic nhiều bước, phân tích nhân quả, kiểm chứng giả thuyết kinh tế/khoa học. |
| 💬 **Core Chat & Advice** | **20%** | Tư vấn tâm lý, đời sống, pháp lý, y tế, mua sắm, giao tiếp đàm thoại tự nhiên chuẩn mực. |
| 💻 **Code & Technical** | **15%** | DevOps, Docker, Kubernetes, CI/CD, thuật toán, lập trình Python, JavaScript, Rust, SQL, gỡ lỗi hệ thống. |
| 🔄 **Multi-Turn Context** | **12%** | Hội thoại nhiều lượt, duy trì nhất quán ngữ cảnh, đào sâu chi tiết, đính chính thông tin. |
| 🛠️ **Tool Use & Agentic** | **10%** | Gọi hàm theo JSON Schema chuẩn, thiết kế quy trình tự động hóa (Workflow), điều phối API. |
| 📚 **Domain Knowledge** | **10%** | Lịch sử Việt Nam, kinh tế vĩ mô, pháp luật Việt Nam, y học phổ thông, văn hóa bản địa. |
| 🛡️ **Safety & Alignment** | **6%** | Từ chối lịch sự và có cơ sở các yêu cầu độc hại, chống thao túng, an toàn bảo mật thông tin. |
| 🎨 **Creative Writing** | **5%** | Sáng tạo nội dung có chiều sâu, kịch bản, email thương mại, tóm tắt và tổng hợp báo cáo. |
---
## 4. Hệ Thống Đánh Giá Chất Lượng 5 Tiêu Chí (G-Eval Rubric)
Mỗi cuộc trò chuyện được đánh giá độc lập theo công thức điểm trọng số:
$$\text{Quality Score} = 2.0 \times \left( 0.20 \times \text{Fluency} + 0.30 \times \text{Factuality} + 0.30 \times \text{Helpfulness} + 0.10 \times \text{Consistency} + 0.10 \times \text{Structure} \right)$$
### Tiêu Chí Chấm Điểm (1.0 – 5.0):
- **Fluency (Độ tự nhiên & Xưng hô - Trọng số 0.20)**: 100% tiếng Việt bản xứ, không dấu vết dịch máy, chuẩn mực xưng hô `mình/tôi - bạn`.
- **Factuality (Độ chính xác - Trọng số 0.30)**: Hoàn toàn chính xác, không hallucinate, chuẩn kiến thức chuyên ngành.
- **Helpfulness (Độ hữu ích & Trọng tâm - Trọng số 0.30)**: Giải quyết trọn vẹn yêu cầu, cô đọng, giàu giá trị thực tế.
- **Consistency (Độ nhất quán logic - Trọng số 0.10)**: Lập luận chặt chẽ, không tự mâu thuẫn qua các lượt trò chuyện.
- **Structure (Bố cục & Định dạng - Trọng số 0.10)**: Định dạng tối ưu theo câu hỏi, không over-format câu đơn giản.
### Phân Hạng Chất Lượng (Quality Gates):
- 🥇 **Gold Tier (≥ 9.0/10.0)**: Hoàn hảo, không tì vết (*100% mẫu preview đều thuộc Gold Tier*).
- 🥈 **Silver Tier (8.0 – 8.9)**: Tốt, dùng bổ trợ cho tập huấn luyện lớn.
- 🥉 **Bronze Tier (7.0 – 7.9)**: Trung bình, lọc bỏ khỏi tập train tinh hoa.
- ❌ **Reject (< 7.0)**: Bị loại bỏ tự động ngay trong quá trình sinh.
---
5. Hướng Dẫn Sử Dụng & Fine-Tuning (Quickstart)
### Tải Dữ Liệu Bằng Hugging Face `datasets`
```python
from datasets import load_dataset
# Tải tập dữ liệu Vietverse-SFT (Bản Preview 100 Gold)
dataset = load_dataset("TTP01/Vietverse-sft-preview", split="train")
print(f"✅ Đã tải: {len(dataset)} mẫu dữ liệu")
print("User:", dataset[0]["messages"][0]["content"])
print("AI:", dataset[0]["messages"][1]["content"])
```
---
### Huấn Luyện LoRA Nhanh Với Unsloth (Qwen 2.5 / Llama 3.1 / DeepSeek)
```python
from unsloth import FastLanguageModel
from trl import SFTTrainer
from transformers import TrainingArguments
from datasets import load_dataset
# 1. Khởi tạo mô hình (4-bit quantization để tiết kiệm VRAM)
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "unsloth/Qwen2.5-7B-Instruct",
max_seq_length = 4096,
load_in_4bit = True,
)
# 2. Cấu hình LoRA Adapters
model = FastLanguageModel.get_peft_model(
model,
r = 16,
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
lora_alpha = 16,
lora_dropout = 0,
bias = "none",
)
# 3. Nạp và định dạng dữ liệu
dataset = load_dataset("TTP01/Vietverse-sft-preview", split="train")
def apply_template(batch):
texts = [tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=False) for msgs in batch["messages"]]
return {"text": texts}
dataset = dataset.map(apply_template, batched=True)
# 4. Chạy huấn luyện
trainer = SFTTrainer(
model = model,
tokenizer = tokenizer,
train_dataset = dataset,
dataset_text_field = "text",
max_seq_length = 4096,
args = TrainingArguments(
per_device_train_batch_size = 2,
gradient_accumulation_steps = 4,
warmup_steps = 10,
max_steps = 200,
learning_rate = 2e-4,
fp16 = True,
logging_steps = 5,
output_dir = "vietverse-qwen-sft",
),
)
trainer.train()
```
---
## 6. Minh Họa Cấu Trúc Dữ Liệu Chi Tiết (Full 35+ Fields Schema)
🔍 Bấm để xem toàn bộ 35+ trường Metadata JSON của mẫu VVSFT-00000001
```json
{
"id": "VVSFT-00000001",
"dataset": "Vietverse-SFT",
"version": "1.0",
"language": "vi",
"messages": [
{
"role": "user",
"content": "Con mình học lớp 2, cô giáo phàn nàn cháu hay ngồi không yên khi học online, mình phân vân giữa khám tâm lý hay chỉ là do con nít hiếu động. Bạn tư vấn giúp mình nên tìm hiểu ở đâu cho hợp lý?"
},
{
"role": "assistant",
"content": "Ngồi không yên và khó tập trung ở tuổi lớp 2 khi học online là phản ứng rất phổ biến. Ở tuổi này, khả năng tập trung chủ động của trẻ duy trì khoảng 15–20 phút. Thay vì vội vàng đưa đi khám tâm lý ngay, bạn có thể kiểm tra một số yếu tố môi trường học tập và các mốc sinh hoạt hàng ngày trước..."
}
],
"task_type": "core_chat",
"subtask": "opinion",
"domain": "education",
"sub_domain": "primary_school",
"topic_tags": [
"education",
"primary_school",
"core_chat",
"opinion"
],
"difficulty": "expert",
"turn_count": 1,
"answer_style": "friendly",
"source_type": "synthetic_llm",
"verified": false,
"region": "vn",
"dialect": "north",
"knowledge_scope": "scope_education_primary_school",
"requires_reasoning": false,
"requires_grounding": false,
"has_tools": false,
"is_multi_turn": false,
"has_code": false,
"quality_score": 9.8,
"fluency_score": 10.0,
"factuality_score": 9.8,
"helpfulness_score": 10.0,
"consistency_score": 10.0,
"structure_score": 10.0,
"human_reviewed": false,
"tier": "gold",
"generation_method": "planner_instruct_output_judge",
"decontam_hash": "sha256:35bccc75f0f76bf2941a378375d264a0",
"near_dup_cluster": "cluster_a704f6193c919462",
"user_persona": "bac_si",
"user_persona_detail": {
"age_group": "30-55",
"education_level": "postgraduate",
"tech_literacy": "intermediate",
"formality_preference": "formal",
"region": "north"
},
"intent": "decide",
"conversation_goal": "explore_topic",
"expected_skills": [
"empathetic_listening",
"legal_interpretation"
],
"reasoning_depth": "deep",
"knowledge_depth": "intermediate",
"pipeline_trace": {
"pipeline_version": "v1.0",
"run_id": "run_20260815_173319",
"batch_id": "batch_core_chat",
"planner_output": {
"task_type": "core_chat",
"subtask": "opinion",
"domain": "education",
"sub_domain": "primary_school",
"difficulty": "expert"
},
"instruction_generation": {
"prompt_template": "vi_stage1_input_gen_v1",
"temperature": 0.85,
"generated_at": "2026-08-15T10:33:57+00:00Z"
},
"output_generation": {
"prompt_template": "vi_stage2_output_gen_v1",
"temperature": 0.7,
"generated_at": "2026-08-15T10:33:57+00:00Z"
},
"judge_output": {
"criteria_reasoning": {
"fluency": "Tiếng Việt tự nhiên, xưng hô mình/bạn nhất quán, không có dấu hiệu dịch máy.",
"factuality": "Thông tin về giai đoạn vàng và tâm lý học đường chính xác, khoa học.",
"helpfulness": "Trả lời trọn vẹn, đưa ra hướng quan sát và lời khuyên cụ thể, đúng trọng tâm.",
"consistency": "Lập luận nhất quán, các luận điểm bổ trợ logic chặt chẽ.",
"structure": "Cấu trúc rõ ràng với các phần và bullet, dễ theo dõi."
},
"fluency": 10.0,
"factuality": 9.8,
"helpfulness": 10.0,
"consistency": 10.0,
"structure": 10.0,
"overall_score": 9.8,
"computed_quality": 9.8,
"verdict": "PASS",
"judge_notes": "Phản hồi xuất sắc, chuẩn tâm lý giáo dục thực tế, cấu trúc tốt.",
"judged_at": "2026-08-15T10:33:57+00:00Z"
},
"dedup_check": {
"exact_dup": false,
"near_dup": false
}
},
"token_stats": {
"instruction_tokens": 53,
"response_tokens": 599,
"total_tokens": 652,
"response_to_instruction_ratio": 11.3
},
"curriculum_rank": 8074,
"curriculum_bucket": "elite",
"benchmark_safe": true
}
```
---
## 7. Giới Hạn & Đạo Đức Dữ Liệu (Limitations & Ethics)
- **Dữ liệu tổng hợp (Synthetic Data Boundary)**: Dù đã qua 3 tầng lọc và kiểm chứng G-Eval, dữ liệu tổng hợp từ LLM vẫn có thể tồn tại sai sót nhỏ ở các mốc thời gian hoặc số hiệu văn bản pháp lý rất chi tiết. Chúng tôi khuyến nghị người dùng đối chiếu văn bản gốc khi áp dụng cho các quyết định pháp lý và y khoa quan trọng.
- **Tính an toàn**: Tập dữ liệu đã được lọc sạch các nội dung độc hại, vi phạm pháp luật và đảm bảo tính bảo mật cá nhân (No PII).
---
## 8. Ủng Hộ Dự Án & Liên Hệ (Support & Contact)
Nếu bạn thấy dự án **Vietverse-SFT** hữu ích và muốn tiếp thêm nguồn lực (chi phí API & Compute) để mở rộng các tập dữ liệu tiếng Việt quy mô lớn hơn trong tương lai, mọi sự ủng hộ đều rất quý báu:
- 💎 **USDT (Mạng Aptos - APTOS Network)**:
```text
0x9fdd674174f47bbfaeaeaa9b34459c8436b87e20327ebfc7a920d650c76febb2
```
- 📬 **Email liên hệ & Hợp tác**: [`ttprbbull@gmail.com`](mailto:ttprbbull@gmail.com)
---
---
🇬🇧 PART 2: ENGLISH (TECHNICAL DATASET CARD)
> 📢 **PREVIEW / COMMUNITY DEMO NOTICE**
>
> This is the initial **Preview release containing 100 Gold-Tier samples (≥ 9.0/10.0)** from the **Vietverse-SFT** initiative.
> - **Purpose**: To provide the open-source community with an authentic preview of native Vietnamese multi-agent synthesis, conversational nuances, and rich 35-field metadata architecture.
> - ☕ **Full Release Timeline**: *There is currently no fixed timeline for future full-scale releases (further scaling will depend on available funding and spare time :)))).*
---
## 1. Motivation & Context
In the development of Vietnamese Large Language Models (LLMs), the primary barrier to conversational fluency has been **the widespread reliance on machine-translated datasets**. Naive translation of English instruction corpora (such as translated Alpaca, ShareGPT, or UltraChat) introduces severe systematic flaws:
1. **Translation Artifacts & Unnatural Calques**: Questions frequently follow rigid translational formulas (*"Please...", "Do...", "Kindly inform me..."*) that native Vietnamese speakers never use in authentic dialogues.
2. **Cultural & Contextual Disconnect**: Absence of domestic legal, educational, administrative, and cultural context.
3. **Pronoun & Persona Inconsistency**: Unstable and unnatural pronoun switches during multi-turn interactions.
**Vietverse-SFT** is built upon the **Data-Centric AI** paradigm (*Quality over Quantity*), drawing on empirical insights from *LIMA (Zhou et al., 2023)* and *Alpagasus (Chen et al., 2024)*, which prove that a compact, rigorously filtered, and persona-rich instruction corpus outperforms noisy translated corpora of hundreds of thousands of samples.
---
## 2. 3-Stage Multi-Agent Pipeline Architecture
Every conversation in Vietverse-SFT is synthesized and audited through a **3-Stage Multi-Agent Pipeline**:
```
┌────────────────────────────────────────────────────────────────────────────────────────┐
│ VIETVERSE-SFT 3-STAGE MULTI-AGENT PIPELINE │
├──────────────────────────┬──────────────────────────┬──────────────────────────────────┤
│ STAGE 1: USER SIMULATOR │ STAGE 2: DOMAIN EXPERT │ STAGE 3: G-EVAL QUALITY JUDGE │
├──────────────────────────┼──────────────────────────┼──────────────────────────────────┤
│ • 19 Authentic Personas │ • High Information Density│ • 5-Criteria Rubric (1.0 - 5.0) │
│ • Regional Dialects (N/C/S)│ • Zero Robotic AI-tell │ • Grounded Textual Evidence │
│ • Evol-Instruct Rules │ • Anti-Sycophancy Rebuttal│ • Zero Verbosity/Halo Bias │
│ • Diverse Opening Forms │ • Clean Pronoun Address │ • Automated Quality Gate Gating │
└──────────────────────────┴──────────────────────────┴──────────────────────────────────┘
```
### Methodological Breakdown:
1. **Stage 1 — User Simulator**:
- Built on *Persona Hub (Tencent AI Lab)*: Constructs granular user personas (profession, familiarity level, psychological state, device) prior to prompt generation.
- *Evol-Instruct Complexity Scaling*: Enforces realistic real-world constraints (concrete error messages, software versions, budget limitations, jurisdictional nuances).
- Diverse Openings: Eliminates repetitive prompt starters, supporting direct inquiries, situational context, and prior belief verification.
2. **Stage 2 — Domain Expert Assistant**:
- Generates concise, structured, and knowledge-dense responses.
- **Anti-Sycophancy**: Constructively challenges false user assumptions (*"I heard adding RAM always fixes lag..."*) rather than providing false validation.
- Standardized 1-on-1 human-AI address protocol (`mình/tôi - bạn`).
3. **Stage 3 — G-Eval Quality Judge**:
- Independent LLM Judge evaluates the conversation across 5 weighted dimensions grounded in verifiable textual evidence.
- Strictly penalizes verbosity and superficial authority tone.
---
## 3. Shard Taxonomy & Capability Distribution
### 📊 A. Verified Empirical Preview Breakdown (100 Gold Samples)
All **100 samples in the Preview release** have been empirically verified and categorized across core real-world domains:
```
Education & Pedagogy ████████████████████ 20% (20 samples)
Vietnamese Cultural Knowledge ██████████████████ 18% (18 samples)
Macroeconomics & Finance ████████████████ 16% (16 samples)
Vietnamese Civil & Business Law ████████████████ 16% (16 samples)
Technology, DevOps & Code ███████████████ 15% (15 samples)
General Chat & Daily Advisory ████████ 8% (8 samples)
Healthcare & Clinical Insights ███████ 7% (7 samples)
```
| Statistical Metric | Empirical Value in 100-Sample Preview |
| :--- | :--- |
| **🥇 Gold Tier Ratio (≥ 9.0/10.0)** | **100% (100/100 Gold-Tier verified)** |
| **⭐ Mean Quality Score** | **`9.63 / 10.0`** (Range: `9.0` – `10.0`) |
| **📏 Mean Sample Length** | **`928.9 tokens / sample`** |
| **🗺️ Regional Dialect Distribution** | North (`north`): **49%** • South (`south`): **48%** • Central (`central`): **3%** |
| **🎯 Difficulty Granularity** | `expert`: **29%** • `easy`: **29%** • `hard`: **21%** • `medium`: **21%** |
| **🔄 Turn Count Distribution** | 1-turn: **90 samples** • 2-turn: **4 samples** • 3-turn: **3 samples** • 4-turn: **3 samples** |
---
### 🗺️ B. Target Design Taxonomy Overview
```
Reasoning & Logic ██████████████████████ 22%
Core Chat & Advice ████████████████████ 20%
Code & Technical ███████████████ 15%
Multi-Turn Dialogue ████████████ 12%
Tool Use & Agentic ██████████ 10%
Domain Knowledge ██████████ 10%
Safety & Alignment ██████ 6%
Creative Writing █████ 5%
```
| Shard / Category | Ratio | Core Capabilities & Focus Domains |
| :--- | :---: | :--- |
| 🧠 **Reasoning & Logic** | **22%** | Multi-step mathematical reasoning, formal logic, cause-effect deduction, scientific hypothesis verification. |
| 💬 **Core Chat & Advice** | **20%** | Everyday consultation, psychology, civil law, healthcare, natural conversational interactions. |
| 💻 **Code & Technical** | **15%** | DevOps, Docker/Kubernetes, CI/CD pipelines, algorithms, Python, JS, Rust, SQL, systems debugging. |
| 🔄 **Multi-Turn Context** | **12%** | Multi-turn dialogue, context repair, progressive refinement, clarification, debate. |
| 🛠️ **Tool Use & Agentic** | **10%** | JSON Schema function calling, autonomous workflow design, API orchestration. |
| 📚 **Domain Knowledge** | **10%** | Vietnamese history, macroeconomics, civil jurisprudence, local cultural context, medicine. |
| 🛡️ **Safety & Alignment** | **6%** | Grounded refusal of malicious queries, anti-sycophancy, information security. |
| 🎨 **Creative Writing** | **5%** | High-density content synthesis, video scriptwriting, commercial communications, executive summaries. |
---
## 4. 5-Criteria Quality Evaluation Rubric (G-Eval)
Each sample is scored on a **1.0 – 5.0 scale** across 5 weighted criteria (scaled to 10.0):
$$\text{Quality Score} = 2.0 \times \left( 0.20 \times \text{Fluency} + 0.30 \times \text{Factuality} + 0.30 \times \text{Helpfulness} + 0.10 \times \text{Consistency} + 0.10 \times \text{Structure} \right)$$
### Evaluation Dimensions:
- **Fluency (Weight: 0.20)**: 100% natural native Vietnamese, zero translation artifacts, standard 1-on-1 addressing.
- **Factuality (Weight: 0.30)**: Verifiably accurate, zero hallucinations, domain-grounded.
- **Helpfulness (Weight: 0.30)**: Directly answers the user's intent with high information density.
- **Consistency (Weight: 0.10)**: Strict logical coherence across turns.
- **Structure (Weight: 0.10)**: Clean formatting calibrated to prompt complexity without excessive markdown bloat.
### Quality Gating:
- 🥇 **Gold Tier (≥ 9.0/10.0)**: Flawless, verified accuracy, native fluency (*100% of preview samples are Gold Tier*).
- 🥈 **Silver Tier (8.0 – 8.9)**: High quality, suitable for secondary large-scale training.
- 🥉 **Bronze Tier (7.0 – 7.9)**: Average quality, excluded from elite training runs.
- ❌ **Reject (< 7.0)**: Instantly pruned during pipeline execution.
---
## 5. Quickstart & Fine-Tuning
### Loading with Hugging Face `datasets`
```python
from datasets import load_dataset
# Load Vietverse-SFT (100 Gold Preview Edition)
dataset = load_dataset("TTP01/Vietverse-sft-preview", split="train")
print(f"✅ Loaded {len(dataset)} samples")
print("User:", dataset[0]["messages"][0]["content"])
print("Assistant:", dataset[0]["messages"][1]["content"])
```
---
### Fast Fine-Tuning with Unsloth
```python
from unsloth import FastLanguageModel
from trl import SFTTrainer
from transformers import TrainingArguments
from datasets import load_dataset
# 1. Load Base Model with 4-bit Quantization
model, tokenizer = FastLanguageModel.from_pretrained(
model_name = "unsloth/Qwen2.5-7B-Instruct",
max_seq_length = 4096,
load_in_4bit = True,
)
# 2. Configure LoRA Adapters
model = FastLanguageModel.get_peft_model(
model,
r = 16,
target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
lora_alpha = 16,
lora_dropout = 0,
bias = "none",
)
# 3. Format Chat Template
dataset = load_dataset("TTP01/Vietverse-sft-preview", split="train")
def format_prompts(batch):
texts = [tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=False) for msgs in batch["messages"]]
return {"text": texts}
dataset = dataset.map(format_prompts, batched=True)
# 4. Train Model
trainer = SFTTrainer(
model = model,
tokenizer = tokenizer,
train_dataset = dataset,
dataset_text_field = "text",
max_seq_length = 4096,
args = TrainingArguments(
per_device_train_batch_size = 2,
gradient_accumulation_steps = 4,
warmup_steps = 10,
max_steps = 200,
learning_rate = 2e-4,
fp16 = True,
logging_steps = 5,
output_dir = "vietverse-qwen-sft",
),
)
trainer.train()
```
---
## 6. Complete Data Schema Specification (Full 35+ Fields Schema)
🔍 Click to expand full 35+ fields Metadata JSON of instance VVSFT-00000001
```json
{
"id": "VVSFT-00000001",
"dataset": "Vietverse-SFT",
"version": "1.0",
"language": "vi",
"messages": [
{
"role": "user",
"content": "Con mình học lớp 2, cô giáo phàn nàn cháu hay ngồi không yên khi học online, mình phân vân giữa khám tâm lý hay chỉ là do con nít hiếu động. Bạn tư vấn giúp mình nên tìm hiểu ở đâu cho hợp lý?"
},
{
"role": "assistant",
"content": "Ngồi không yên và khó tập trung ở tuổi lớp 2 khi học online là phản ứng rất phổ biến. Ở tuổi này, khả năng tập trung chủ động của trẻ duy trì khoảng 15–20 phút. Thay vì vội vàng đưa đi khám tâm lý ngay, bạn có thể kiểm tra một số yếu tố môi trường học tập và các mốc sinh hoạt hàng ngày trước..."
}
],
"task_type": "core_chat",
"subtask": "opinion",
"domain": "education",
"sub_domain": "primary_school",
"topic_tags": [
"education",
"primary_school",
"core_chat",
"opinion"
],
"difficulty": "expert",
"turn_count": 1,
"answer_style": "friendly",
"source_type": "synthetic_llm",
"verified": false,
"region": "vn",
"dialect": "north",
"knowledge_scope": "scope_education_primary_school",
"requires_reasoning": false,
"requires_grounding": false,
"has_tools": false,
"is_multi_turn": false,
"has_code": false,
"quality_score": 9.8,
"fluency_score": 10.0,
"factuality_score": 9.8,
"helpfulness_score": 10.0,
"consistency_score": 10.0,
"structure_score": 10.0,
"human_reviewed": false,
"tier": "gold",
"generation_method": "planner_instruct_output_judge",
"decontam_hash": "sha256:35bccc75f0f76bf2941a378375d264a0",
"near_dup_cluster": "cluster_a704f6193c919462",
"user_persona": "bac_si",
"user_persona_detail": {
"age_group": "30-55",
"education_level": "postgraduate",
"tech_literacy": "intermediate",
"formality_preference": "formal",
"region": "north"
},
"intent": "decide",
"conversation_goal": "explore_topic",
"expected_skills": [
"empathetic_listening",
"legal_interpretation"
],
"reasoning_depth": "deep",
"knowledge_depth": "intermediate",
"pipeline_trace": {
"pipeline_version": "v1.0",
"run_id": "run_20260815_173319",
"batch_id": "batch_core_chat",
"planner_output": {
"task_type": "core_chat",
"subtask": "opinion",
"domain": "education",
"sub_domain": "primary_school",
"difficulty": "expert"
},
"instruction_generation": {
"prompt_template": "vi_stage1_input_gen_v1",
"temperature": 0.85,
"generated_at": "2026-08-15T10:33:57+00:00Z"
},
"output_generation": {
"prompt_template": "vi_stage2_output_gen_v1",
"temperature": 0.7,
"generated_at": "2026-08-15T10:33:57+00:00Z"
},
"judge_output": {
"criteria_reasoning": {
"fluency": "Tiếng Việt tự nhiên, xưng hô mình/bạn nhất quán, không có dấu hiệu dịch máy.",
"factuality": "Thông tin về giai đoạn vàng và tâm lý học đường chính xác, khoa học.",
"helpfulness": "Trả lời trọn vẹn, đưa ra hướng quan sát và lời khuyên cụ thể, đúng trọng tâm.",
"consistency": "Lập luận nhất quán, các luận điểm bổ trợ logic chặt chẽ.",
"structure": "Cấu trúc rõ ràng với các phần và bullet, dễ theo dõi."
},
"fluency": 10.0,
"factuality": 9.8,
"helpfulness": 10.0,
"consistency": 10.0,
"structure": 10.0,
"overall_score": 9.8,
"computed_quality": 9.8,
"verdict": "PASS",
"judge_notes": "Phản hồi xuất sắc, chuẩn tâm lý giáo dục thực tế, cấu trúc tốt.",
"judged_at": "2026-08-15T10:33:57+00:00Z"
},
"dedup_check": {
"exact_dup": false,
"near_dup": false
}
},
"token_stats": {
"instruction_tokens": 53,
"response_tokens": 599,
"total_tokens": 652,
"response_to_instruction_ratio": 11.3
},
"curriculum_rank": 8074,
"curriculum_bucket": "elite",
"benchmark_safe": true
}
```
---
## 7. Limitations & Ethical Considerations
- **Synthetic Data Boundary**: While audited through 3-stage multi-agent verification, synthetic generations may occasionally contain minor factual inaccuracies in niche legal citations or dates. Users should cross-reference original primary sources for critical legal or medical decisions.
- **Safety**: The corpus has been decontaminated against personally identifiable information (PII) and toxic patterns.
---
## 8. License & Citation
- **License**: Released under the permissive **Apache License 2.0**.
- **Citation**:
```bibtex
@dataset{vietverse_sft_preview_2026,
author = {TTP01 and Vietverse Contributors},
title = {Vietverse-SFT: A High-Density, Native Vietnamese Foundation Dataset for Large Language Model Instruction Tuning},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/TTP01/Vietverse-sft-preview}},
version = {1.0-preview}
}
```
---
## 9. Support the Project & Contact
If you find **Vietverse-SFT** valuable and wish to support future API compute expenses for scaling up upcoming full releases, donations are warmly appreciated:
- 💎 **USDT (Aptos Network)**:
```text
0x9fdd674174f47bbfaeaeaa9b34459c8436b87e20327ebfc7a920d650c76febb2
```
- 📬 **Contact & Inquiries**: [`ttprbbull@gmail.com`](mailto:ttprbbull@gmail.com)
---
Built with ❤️ for the Vietnamese AI & Open-Source Community.