BunGPT-DeepEdu-Decision-4B-Open
A model for fast product decisions without text generation
2 min read
The model evaluates a state (message, dialogue, or document) and multiple choices in a single pass, outputting probabilities and raw scores for each option. It is built on Qwen3.5-4B with an added custom scoring head. Unlike standard LLMs (large language models), this tool does not generate text; instead, it returns structured responses for classification and routing. Developers can integrate the model into products for message filtering, intent recognition, or query routing.
Key features
- Single-pass option evaluation: Processes a document and multiple questions with multiple-choice or yes/no options simultaneously.
- High execution speed: Shares the common request context, speeding up the processing of multiple questions.
- Out-of-the-box HTTP server: Supports an API compatible with the Jev Decision Index format.
- Open-source inference: The Python package and server are distributed under the Apache-2.0 license.
What you can do
- Integrate into a chatbot: Set up automated user intent recognition for query routing.
- Add filtering: Check messages for safety and compliance with school environment rules.
- Analyze documents: Evaluate text compliance against given criteria or classify request types.
Example
from deepedu_decision import DecisionModel
model = DecisionModel.from_pretrained("DeepEduVN/bungpt-open", device="cuda:0")
state = "Học sinh: Em không hiểu vì sao đạo hàm của x^2 là 2x..."
questions = {
"intent": {
"type": "choice",
"instructions": "What does the student want?",
"criteria": {"explain": "An explanation", "solve": "A solution"},
}
}
print(model.decide(state, questions))
Read more
DetailsBenchmarks: On the DeepEdu benchmark, the average score is 0.702 (best in routing is 0.820).
Hardware requirements: Tested on an RTX 3090 GPU, prefix caching is supported.
Limitations: Weaker in Vietnamese law and knowledge questions compared to competitors; evaluates only a fixed set of options and is not a chat model.
Architecture: Qwen3.5-4B (4.66 billion parameters) with an additional scoring head.










Comments