MCINext/synthetic-persian-chatbot-rag-faq-pair-classification

Dataset Summary

Synthetic Persian Chatbot RAG FAQ Pair Classification (SynPerChatbotRAGFAQPC) is a Persian (Farsi) dataset for the Pair Classification task, specifically designed for evaluating Retrieval-Augmented Generation (RAG) chatbot systems. It is part of the FaMTEB (Farsi Massive Text Embedding Benchmark) and was synthetically generated using GPT-4o-mini.

The dataset measures a model’s ability to assess whether a given FAQ (question–answer pair) is relevant to a new user message within the context of an ongoing conversation.

• Language(s): Persian (Farsi)
• Task(s): Pair Classification (FAQ Relevance in RAG Chatbots)
• Source: Synthetically generated using GPT-4o-mini, derived from the Synthetic Persian Chatbot RAG Dataset
• Part of FaMTEB: Yes

Supported Tasks and Leaderboards

The dataset evaluates the capability of embedding models to semantically classify FAQ responses as relevant or irrelevant to user queries in a RAG setting. Model results can be found on the Persian MTEB Leaderboard under the Pair Classification category.

Construction

1. GPT-4o-mini generated RAG-style chatbot conversations including multi-turn histories.
2. For each user message, two FAQ pairs were created:
   • Positive: Directly addresses the user's intent.
   • Negative: Topically related but not a suitable response.
3. Each data point is a triplet:
   • (conversation + user message, FAQ pair, label)
   • Label is 1 for positive match, 0 for negative.
4. Human evaluators verified a subset of examples, with 85.5% accuracy achieved.

Data Splits

• Train: 6,522 samples
• Development (Dev): 0 samples
• Test: 2,174 samples