SupraLabs/chat-titles-12K

12K Chat Titles • Filtered Samples

This Dataset is outdated!!!!

This dataset is outdated. please use our newer datasets:

• Filtered
• Unfiltered

About Dataset

Supra Titles 12K is a curated chat title generation dataset derived from the training pipeline used for the experimental Supra Title model family. The dataset is designed specifically for training, fine-tuning, and evaluating models that generate concise, descriptive titles from a user's first message in a conversation.

Unlike general instruction-following datasets, Supra Titles 12K focuses on a single task: transforming a user message into a high-quality chat title that accurately captures the intent, topic, or context of the conversation while remaining concise and readable.

Dataset Creation Process

The dataset was constructed from filtered conversation data originating from the sam-mosaic/orca-gpt4-chatml dataset. User messages were extracted, cleaned, and processed through a multi-stage quality pipeline designed to improve consistency and reduce noise.

The generation process included:

1. Data Extraction

   • User messages were extracted from the source dataset.
   • Only relevant user-facing content was retained.
   • System prompts, assistant responses, metadata, and unnecessary formatting were removed where appropriate.

2. Title Generation

   • High-quality titles were generated using Qwen3.6-35B-A3B.
   • The model was instructed to create concise and descriptive titles that reflect the primary intent of each user message.
   • Titles were generated with a focus on readability, relevance, and practical usage in real-world chat applications.

3. Cleaning and Deduplication

   • Duplicate samples were removed.
   • Low-quality, malformed, or ambiguous entries were filtered out.
   • Formatting inconsistencies were corrected to improve overall dataset quality.

4. Quality Review

   • Additional review and validation passes were performed using Qwen3.5-9B-Q8.
   • Samples failing quality standards were removed or regenerated.
   • The resulting dataset was manually inspected throughout development to ensure consistency across the final release.

Dataset Structure

Each dataset entry contains two fields:

Field	Description
user	The processed user message extracted from the source dataset.
title	A generated title describing the user's message.

Example:

{
  "user": "I'm learning about vector embeddings and semantic search systems, and I keep seeing cosine similarity being used as the primary method for comparing embeddings. Can you explain how cosine similarity actually works mathematically, why it measures similarity based on the angle between vectors instead of their magnitude, and why it is so commonly used in modern embedding models for tasks such as semantic search, document retrieval, recommendation systems, and clustering?",
  "title": "Cosine Similarity Explained"
}

Attribution

If you are using this dataset, kindly attribute SupraLabs:

@dataset{supralabs_chat_titles_12k_2026,
  author       = {QyrouNnet-AI},
  title        = {Chat Titles 12K},
  year         = {2026},
  organization = {SupraLabs},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/SupraLabs/chat-titles-12K}
}