Dataset Card for SID (Synthetic Intent Dataset)

SID (Synthetic Intent Dataset) is a high-quality, augmented multilingual dataset designed for training robust intent classifiers for customer service chatbots. It covers 101 distinct intents across 11 categories and 5 languages.

Dataset Summary

The dataset was curated and augmented to address common challenges in production NLU systems:

• Typo Robustness: Includes simulated physical keyboard errors (QWERTY proximity).
• Grammatical Diversity: Includes simulated grammatical errors across all supported languages.
• Entity Volatility: Features resampled entities (dates, IDs, values) to prevent overfitting.
• Domain Specificity: Optimized for customer support, billing, and technical assistance.

Supported Languages

• English (en)
• Portuguese (pt)
• Spanish (es)
• German (de)
• French (fr)

Dataset Details

• Curated by: @Luigicfilho
• Funded by: @Luigicfilho
• Shared by: @Luigicfilho
• Language(s) (NLP): English (en), Portuguese (pt), Spanish (es), German (de), French (fr)
• License: MIT

Dataset Sources

• Repository: https://huggingface.co/datasets/Luigicfilho/sid

Uses

Direct Use

This dataset is intended for training and benchmarking NLU (Natural Language Understanding) models, specifically Intent Classifiers for multilingual customer service applications. It is optimized for robustness against typos and grammatical errors.

Out-of-Scope Use

• Use in life-critical systems where incorrect intent classification could lead to physical harm.
• Use with languages not included in the metadata.
• Fine-tuning for tasks other than text classification (e.g., generation).

Dataset Structure

Data Fields

• text: The input utterance (string).
• label: The target intent (101 classes, string).
• language: ISO code of the utterance's language (string).

Data Splits

Split	Rows	Description
train	170,511	Augmented training set with typos and grammar simulation.
validation	7,725	Synthetic validation set for model selection.
validation_real	5,201	High-priority real-world human-verified evaluation set.
test	7,725	Synthetic test set for final performance reporting.

Intent Categories

The dataset includes 101 intents organized into the following categories:

• greetings_and_social
• account_and_profile
• orders_and_purchases
• billing_and_payments
• product_and_service_info
• technical_support
• appointments_and_scheduling
• complaints_and_feedback
• information_and_navigation
• legal_and_compliance
• miscellaneous

Usage Example

The easiest way to start using the dataset is with pandas and scikit-learn. Here is a simple baseline for training:

import pandas as pd
from sklearn.preprocessing import LabelEncoder
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import SGDClassifier

# 1. Load the remediated training data
df = pd.read_csv("data/train.csv") 

# 2. Encode labels
le = LabelEncoder()
df['label_id'] = le.fit_transform(df['label'].astype(str))

# 3. Quick Vectorization (TF-IDF)
vectorizer = TfidfVectorizer(ngram_range=(1, 2), max_features=20000)
X = vectorizer.fit_transform(df['text'].astype(str))

# 4. Train a fast linear classifier
clf = SGDClassifier(loss="log_loss")
clf.fit(X, df['label_id'])

# Example Prediction
sample = ["I need to cancel my order please"]
vec_sample = vectorizer.transform(sample)
print(f"Predicted: {le.inverse_transform(clf.predict(vec_sample))[0]}")

Dataset Creation

Curation Rationale

Real-world chatbot users frequently make typing mistakes and grammatical errors. Most public datasets are "too clean," leading to model failure in production. SID was created to provide a specialized benchmark for Typo-Robustness and Grammatical Variance.

Source Data

Data Collection and Processing

The dataset was initially generated using synthetic seed templates for each intent. It was then processed through an automated augmentation pipeline:

1. Typo Simulation: Using QWERTY-proximity character substitution.
2. Grammar Simulation: Injecting common grammatical errors (verb conjugation, gender agreement) for all 5 languages.
3. Entity Resampling: Replacing dates, prices, and IDs with random values to prevent overfitting.
4. UTF-8 Normalization: Ensuring consistent handling of accents and special characters.

Who are the source data producers?

Generated and augmented by @Luigicfilho using specialized NLU augmentation engines.

Annotations

Annotation process

Automated generation based on predefined intent mapping in intents.json.

Who are the annotators?

Automated script with manual spot-checking for the validation_real set.

Personal and Sensitive Information

None. All Names, Order IDs, and Personal Data are synthetically generated or resampled.

Bias, Risks, and Limitations

• The dataset is synthetic; while it simulates errors, it may not capture 100% of the nuance found in natural human-to-human conversations.
• Some niche cultural idioms might be underrepresented in the 5 languages.

Recommendations

Users should be made aware that this dataset is heavily augmented. It is recommended to evaluate models on the validation_real split for the most accurate measure of production-readiness.

Citation

BibTeX:

@dataset{sid_synthetic_intent_2026,
  author = {Luigicfilho},
  title = {SID: Synthetic Intent Dataset},
  year = {2026},
  url = {https://huggingface.co/datasets/Luigicfilho/sid}
}

APA:

Luigicfilho. (2026). SID: Synthetic Intent Dataset (1.0) [Data set]. Hugging Face. https://huggingface.co/datasets/Luigicfilho/sid

Dataset Card Contact

@Luigicfilho