Overview

The agentlans/Qwen3-0.6B-proposition-extractor is a specialized 0.8 billion parameter model built on the Qwen3 architecture. Its primary function is to extract distinct propositions from text paragraphs, presenting them in an HTML-like <ul><li> format. This model is particularly useful for tasks requiring the breakdown of complex text into individual, digestible statements.

Key Capabilities

• Proposition Extraction: Identifies and extracts core propositions from input text.
• HTML-like I/O: Designed to handle input and generate output using HTML paragraph (<p>) and unordered list (<ul><li>) tags, facilitating integration with web-based applications or structured data pipelines.
• Configurable Sampling: Recommends warmer temperatures for sampling to achieve a balance between literal extraction and paraphrasing, avoiding mere sentence listing.

Training and Limitations

The model was trained using LLaMA Factory over 3 epochs, incorporating flashattn2, NEFTune alpha 5, and LoRA with rank 16 and alpha 32. While effective, the quality of results is influenced by sampling parameters and input quality, as natural language can be imprecise. As a large language model, its output is not deterministic.

Good For

• Structured Information Extraction: Ideal for converting unstructured text into a list of clear, distinct propositions.
• Content Summarization: Can be used as a component in systems that aim to summarize documents by highlighting key statements.
• Data Preprocessing: Useful for preparing text data for further analysis where individual propositions are required.