Dingy-Kang/DeFT-RAG-DM-4B

DeFT-RAG Decision Model (4B)

The fine-tuned decision model (DM) of DeFT-RAG, from the paper DeFT-RAG: Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers (Dingyi Kang and Bingzhe Li, University of Texas at Dallas). Code: https://github.com/Dingyi-Kang/DeFT-RAG

This is the model behind every DeFT-RAG result in the paper, with both backbones (gpt-4o-mini and Qwen3.6-35B-A3B).

What it does

DeFT-RAG answers a multi-hop question by asking the LLM for one self-contained sub-question at a time. Every decision in that process is made by this model: it returns a probability for each yes/no question, and a score for the question depth.

• Relevance: is a retrieved passage needed to answer the (sub-)question? (kept if p(yes) ≥ 0.5)
• Sufficiency: do the question's top-10 passages contain everything needed to answer it? (p(yes) ≥ 0.7)
• Question depth: how many hops (chained lookups) does the original question need: none, one, two, or three or more? It is estimated from the question alone, as a score from 0 to 3.
• Direct answer: the question is answered directly from its passages when they are sufficient and its estimated depth is at most 2; otherwise it goes to the loop of sub-questions.
• Enough facts: do the question-answer pairs found so far already answer the original question? (≥ 0.5)
• Same question (QA cache only): do two sub-questions ask for the same information? (≥ 0.8)

The exact request for each decision is in deftrag/loop.py of the code repository.

Model

• Base model: StartLux-Decision-4B, a language model that answers typed questions about a state with a probability for each option.
• Fine-tuning: LoRA (rank 16, alpha 32, dropout 0.05) on all attention and MLP projections, 1 epoch, peak learning rate 1e-4, micro-batches of 8,192 tokens, with the training scripts of the StartLux-Decision release; then one temperature per question type fitted on the development examples (decision_config.json).
• Data: 48,036 training and 7,806 development examples, labelled only from the gold annotations of the training splits of MuSiQue, 2WikiMultiHopQA and HotpotQA. No test question, test passage or test-corpus paragraph was used. The scripts that build the data are in decision_model/ of the code repository.

Files

• The repository root holds the merged model (base weights + LoRA), which is the model served for the paper.
• adapter/ holds the LoRA adapter alone (adapter_model.safetensors, adapter_config.json) and the calibration file, for use on top of StartLux-Decision-4B.

Usage

The model is served with the server of the StartLux-Decision release, which ships with the base model:

huggingface-cli download startlux-models/StartLux-Decision-4B --local-dir StartLux-Decision-4B   # the server code
huggingface-cli download Dingy-Kang/DeFT-RAG-DM-4B --local-dir DeFT-RAG-DM-4B                   # this model
cd StartLux-Decision-4B && pip install -r requirements.txt
python -m startlux_decision.server --model ../DeFT-RAG-DM-4B --port 8090 --no-images --max-length 16384 --name deftrag-dm

DeFT-RAG then reaches it at http://localhost:8090/v1/systemone (its default). One GPU with about 20 GB of free memory is enough. See the code repository for running and grading DeFT-RAG.

License

CC BY-NC 4.0, the license of StartLux-Decision-4B, from which this model is derived: research and other non-commercial use only, with attribution. See LICENSE and NOTICE.

Citation

If you use this model, please cite the DeFT-RAG paper and StartLux-Decision:

@article{kang2026deftrag,
  title   = {{DeFT-RAG}: Decomposing Multi-Hop Questions into Self-Contained Sub-Questions with Reusable Answers},
  author  = {Kang, Dingyi and Li, Bingzhe},
  journal = {arXiv preprint arXiv:2610.37826},
  year    = {2026}
}

@misc{startlux2026,
  title        = {{StartLux-Decision-4B}},
  author       = {{StartLux Labs}},
  howpublished = {\url{https://huggingface.co/startlux-models/StartLux-Decision-4B}},
  year         = {2026}
}