bratao/Qwen3OIE-4B

Qwen3OIE-4B: Portuguese Abstractive Open Information Extraction

Qwen3OIE-4B is a 4-billion parameter decoder-only causal language model developed by bratao, fine-tuned from the Qwen/Qwen3-4B base model. Its primary function is abstractive Open Information Extraction (OpenIE) for the Portuguese language, generating structured binary extractions in JSON format (ARG0, V, ARG1).

Key Capabilities & Features

• Specialized Portuguese OpenIE: Designed specifically for extracting information from Portuguese text, producing structured triples.
• Abstractive Extraction: Generates extractions that may not be literal spans from the source text, offering flexibility in output.
• JSON Output: Provides extractions in a machine-readable JSON format, facilitating integration into downstream applications.
• Performance: Achieved a perfect-match F1 score of 0.3319 on a silver-standard evaluation set of 100 Portuguese sentences, outperforming other models in its evaluation.
• Ease of Use: Can be easily integrated using the portuguese-openie library, which handles model downloading and inference.

Training and Limitations

The model was fine-tuned on 29,026 Portuguese training sentences and 102,788 synthetic OpenIE extractions generated from Wikipedia paragraphs using Gemini 2.5 Flash. It requires approximately 10-12 GB of VRAM for inference. Users should be aware that generated triples can sometimes be incomplete, duplicated, hallucinated, or malformed. The evaluation set is small and primarily encyclopedic, so generalization to diverse Portuguese text types (conversational, dialectal, specialized) is not guaranteed. Extracted claims are not fact-verified and should not be used for high-impact decisions without further validation.