nhminh107/VietEmbed-RAG-Science

VietEmbed-RAG Science

VietEmbed-RAG Science is a Vietnamese retrieval dataset containing 68,567 query-document examples across seven scientific and technical domains.

Each record consists of:

• A Vietnamese query (anchor)
• A relevant passage (positive)
• A semantically related but non-answering passage (hard_negative)
• Topic and domain metadata

The dataset is designed for training and domain adaptation of Vietnamese text embedding, semantic retrieval, and Retrieval-Augmented Generation (RAG) models.

Dataset Overview

Property	Value
Language	Vietnamese
Number of records	68,567
Format	JSON Lines
Split	Train
Domains	7
File size	Approximately 71.4 MB
Positive passages	Yes
Hard-negative passages	Yes

Domain Distribution

Domain	Records
Earth Science	11,484
Biology	10,963
Information Technology	10,612
Astronomy	10,569
Physics	9,014
Chemistry	8,623
Math	7,302
Total	68,567

Dataset Schema

Each JSON object contains the following fields:

Field	Type	Description
domain	string	High-level scientific or technical domain
topic	string	More specific subject within the domain
title	string	Short title describing the document
anchor	string	Query used for retrieval
positive	string	Relevant passage that answers the query
hard_negative	string	Related passage that does not answer the query directly

Example:

{
  "domain": "Physics",
  "topic": "cơ học lượng tử",
  "title": "Nguyên lý bất định Heisenberg",
  "anchor": "Nguyên lý bất định Heisenberg mô tả mối quan hệ giữa những đại lượng nào?",
  "positive": "Nguyên lý bất định Heisenberg phát biểu rằng vị trí và động lượng của một hạt không thể đồng thời được xác định với độ chính xác tùy ý.",
  "hard_negative": "Cơ học lượng tử mô tả hành vi của vật chất và năng lượng ở thang nguyên tử và hạ nguyên tử bằng hàm sóng và các toán tử."
}

Loading the Dataset

Install the Hugging Face datasets library:

python -m pip install datasets

Load the dataset directly from the Hub:

from datasets import load_dataset

dataset = load_dataset(
    "nhminh107/VietEmbed-RAG-Science",
    split="train",
)

print(dataset)
print(dataset[0])

Using the Dataset for Retrieval Training

The fields can be mapped to a standard retrieval triplet:

query = record["anchor"]
positive_document = record["positive"]
negative_document = record["hard_negative"]

A typical contrastive-learning objective should increase the similarity between anchor and positive, while decreasing the similarity between anchor and hard_negative.

Possible applications include:

• Vietnamese bi-encoder and embedding-model training
• Scientific-domain retrieval
• Hard-negative contrastive learning
• Semantic search
• Retrieval-Augmented Generation
• Domain adaptation for Vietnamese information retrieval

Data Construction

The dataset was created using an LLM-assisted Vietnamese document-generation pipeline.

Post-processing included:

1. JSON and schema validation
2. Unicode and whitespace normalization
3. Field-length validation
4. Exact duplicate removal across queries and passages
5. Positive-negative overlap filtering
6. Topic and domain normalization
7. Per-domain semantic deduplication using character n-gram TF-IDF similarity

The semantic deduplication threshold was set to 0.95.

From an initial collection of 174,812 records, the processing pipeline retained 68,567 records after validation and deduplication.

Data Quality

The processed dataset contains:

• No malformed JSON records
• No empty required fields
• No exact duplicate queries
• No exact duplicate positive passages
• No exact duplicate hard-negative passages
• No records where positive and hard_negative are identical

Hard negatives are intended to remain scientifically related to the query while not directly providing its answer.

Limitations

This is a synthetically generated dataset. Although generation and filtering were constrained toward scientific content, individual records have not all been manually reviewed by domain experts.

Potential limitations include:

• Factual inaccuracies or oversimplifications
• Ambiguous questions
• Hard negatives that may partially imply the answer
• Uneven difficulty across domains and topics
• Style patterns introduced by the generation model
• Lack of predefined validation and test splits

Users should perform additional manual review or domain-specific validation before using the dataset for high-stakes scientific, educational, medical, or safety-critical applications.

This dataset should not be treated as an authoritative scientific reference or as a standalone benchmark.

Recommended Evaluation Practice

Because the repository currently provides only a training split, users should create their own validation and test partitions before evaluating a model.

When splitting the dataset, consider grouping semantically related queries to reduce information leakage between train and evaluation sets.

Citation

If you use this dataset, please cite the Hugging Face repository:

@misc{vietembed_rag_science,
  author       = {Ngo Hoang Minh},
  title        = {VietEmbed-RAG Science},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/nhminh107/VietEmbed-RAG-Science}}
}

Disclaimer

The dataset is provided for research and model-development purposes. Users are responsible for validating the data and determining whether it is suitable for their intended application.