bratao/Qwen3OIE-8B

Overview

bratao/Qwen3OIE-8B is an 8 billion parameter decoder-only causal language model, fine-tuned from Qwen/Qwen3-8B, specifically designed for Portuguese abstractive Open Information Extraction (OpenIE). It generates structured binary extractions in JSON format, comprising ARG0, V, and ARG1 fields from input sentences. This model was developed as part of a doctoral evaluation and achieved the highest lexical-match F1 score among evaluated models.

Key Capabilities & Features

• Portuguese OpenIE: Specializes in extracting relational triples from Portuguese text.
• Abstractive Extraction: The generated fields (ARG0, V, ARG1) are not necessarily literal spans of the source sentence.
• JSON Output: Provides structured output, facilitating programmatic use and integration.
• High Lexical-Match F1: Achieved an F1 score of 0.5612 on a silver-standard test set for lexical match, indicating strong performance in token overlap for extractions.
• Integration: Designed for use with the portuguese-openie library, simplifying deployment.

Use Cases

This model is ideal for applications requiring structured information extraction from Portuguese natural language. It can be used to:

• Convert unstructured Portuguese text into a structured, machine-readable format.
• Populate knowledge bases or databases with facts extracted from Portuguese documents.
• Support downstream NLP tasks that benefit from explicit subject-verb-object relations in Portuguese.

Limitations

• The test set is small and primarily encyclopedic; performance on diverse Portuguese text (dialectal, conversational, specialized) is unknown.
• Generative output may occasionally omit, duplicate, or hallucinate relations, or violate the requested JSON schema.
• The training status suggests it may be a research checkpoint, and users should pin revisions for reproducible work.