olaverse/mist-qg-1.5b

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 2, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The olaverse/mist-qg-1.5b is a compact 1.5 billion parameter multilingual question generation model developed by Olaverse, built on the Qwen/Qwen2.5-1.5B-Instruct backbone. It is specifically designed to produce natural, search-style questions from a given passage across 25 languages, with a maximum sequence length of 3072 tokens. This model excels at generating questions that are directly answerable by the provided text, making it suitable for creating training data for retrievers and rerankers, and for generating questions for various applications.

Loading preview...

Overview

olaverse/mist-qg-1.5b is a specialized 1.5 billion parameter language model from Olaverse, engineered for multilingual question generation. Built upon the Qwen/Qwen2.5-1.5B-Instruct backbone, this decoder-only LM is capable of generating structured JSON output containing natural, search-style questions from a given passage. It supports 25 languages, including several African languages often underserved by existing tools.

Key Capabilities

  • Multilingual Question Generation: Generates relevant questions in 25 languages (e.g., English, French, Spanish, Hindi, Japanese, Yoruba, Swahili, Amharic) that are directly answerable by the input passage.
  • Structured JSON Output: Produces questions in a reliable {"questions": ["...", "...", "..."]} format, facilitating integration into automated workflows.
  • Compact and Efficient: At approximately 1.5 billion parameters, it runs efficiently on a single modest GPU.
  • High-Quality Data Generation: Fine-tuned on a distilled dataset where each generated question was verified by round-trip retrieval, ensuring high relevance to the source passage.

Good For

  • Data Factory: Ideal for minting (query, positive) training pairs to fine-tune retriever and reranker models.
  • Search-Style Question Generation: Useful for applications requiring natural, user-query-like questions from text.
  • Multilingual Applications: Particularly valuable for generating questions in a wide array of languages, including those with fewer existing resources.
  • Content Understanding: Can be used to quickly extract potential questions a passage answers, aiding in content analysis or summarization tasks.

Finetunes
1 models