MCINext/synthetic-persian-chatbot-rag-tone-user-classification

Dataset Summary

Synthetic Persian Chatbot RAG Tone User Classification (SynPerChatbotRAGToneUserClassification) is a Persian (Farsi) dataset created for the Classification task, specifically to identify the user's tone in chatbot conversations within Retrieval-Augmented Generation (RAG) contexts. It is part of the FaMTEB (Farsi Massive Text Embedding Benchmark). The dataset was synthetically generated using the GPT-4o-mini Large Language Model and derived from the Synthetic Persian Chatbot RAG Dataset. Each example includes a user-chatbot conversation with varying history lengths, and the goal is to classify the user's conversational tone (e.g., formal, casual, childish, aggressive, street-style).

• Language(s): Persian (Farsi)
• Task(s): Classification (User Tone Classification in RAG Chatbots)
• Source: Synthetic, created using GPT-4o-mini
• Part of FaMTEB: Yes

Supported Tasks and Leaderboards

This dataset is used to evaluate models on their ability to classify user tone in conversational contexts tailored to RAG systems. Model performance is benchmarked on the Persian MTEB Leaderboard on Hugging Face Spaces (filtered by language: Persian).

Construction

The dataset was created by:

• Generating 175 unique chatbot conversation topics
• Simulating user-chatbot dialogues with random conversation history lengths
• Assigning user tones from a set of 19 tone combinations (e.g., formal, casual, childish, aggressive, street-style)
• Using the user’s specified tone as the class label for the classification task

According to the FaMTEB paper (Table 1), the quality of tone labels was evaluated by human annotators, achieving 85% accuracy in human agreement with GPT-4o-mini's tone annotations.

Data Splits

As defined in the FaMTEB paper (Table 5):

• Train: 3,261 samples
• Development (Dev): 0 samples
• Test: 1,087 samples