qwen3-4b-bkee-full-extraction

Qwen3-4B BKEE Full Extraction

An end-to-end Vietnamese event extraction model fine-tuned from unsloth/Qwen3-4B-Instruct-2507. Given a raw sentence, the model jointly extracts entity mentions, event triggers with their types, and argument role assignments in a single pass, together with a step-by-step natural-language reasoning trace covering the full extraction process.

This is one of four task-specific checkpoints from a distant-supervision Vietnamese event-extraction pipeline. The other three isolate each sub-task:

• hadeseus/qwen3-4b-bkee-entity-extraction
• hadeseus/qwen3-4b-bkee-event-extraction
• hadeseus/qwen3-4b-bkee-argument-assigner

Training Data

The model was trained on a silver-labeled reasoning dataset derived from BKEE (a Vietnamese Event Extraction dataset introduced at LREC-COLING 2024, containing 1,066 manually annotated documents across 33+ event types and 28 argument roles).

The silver data was produced via a distant supervision pipeline:

1. Full BKEE gold annotations (entities, triggers, event types, and argument roles) are used to prompt an LLM to generate a single holistic step-by-step Chain-of-Thought reasoning trace that walks through entity identification, event detection, and argument assignment together.
2. Predictions are validated against Pydantic schemas and checked against BKEE gold annotations. Gold labels are treated as absolute truth: when a sentence has no gold entities/events, the LLM is directed to reason about and justify why nothing was found, rather than inventing extractions.
3. Verified reasoning traces are combined with the original BKEE sentences to form the supervised fine-tuning dataset for this checkpoint.

Note: this model is trained on machine-generated (silver) reasoning data aligned to BKEE's gold labels, not on human-written explanations.

Intended Use

Given a Vietnamese sentence, the model outputs a full reasoning trace followed by the complete structured extraction: entities, event triggers/types, and argument role assignments. It is intended as a single-pass alternative to running the entity, event, and argument checkpoints as separate pipeline stages.

Usage

System prompt:

Trích xuất thực thể kèm span text, trigger sự kiện và tham số tương ứng của từng thực thể từ câu tiếng Việt đã cho. Hãy suy nghĩ theo từng bước.

User prompt: the raw sentence. Output: reasoning followed by {"entities": [...], "events": [{"type": ..., "trigger": ..., "arguments": [...]}]}.

Citation

If you use this model, please cite the underlying BKEE dataset:

@inproceedings{nguyen-etal-2024-bkee,
    title = "{BKEE}: Pioneering Event Extraction in the {V}ietnamese Language",
    author = "Nguyen, Thi-Nhung  and
      Tran, Bang Tien  and
      Luu, Trong-Nghia  and
      Nguyen, Thien Huu  and
      Nguyen, Kiem-Hieu",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.217",
    pages = "2421--2427"
}