DCEA-RAG

DCEA-RAG v18: code, checkpoints and reproducibility artifacts

Ziqi Zhong, Jinhai Huang and Jian An. School of Computer Science and Technology, Xi'an Jiaotong University. Correspondence: [email protected].

DCEA-RAG assembles ordered evidence under a strict unique-passage budget. The frozen scorer supplies hop-specific unary support and directed prerequisite compatibility. This repository distributes the v18 code, two frozen checkpoints and the primary evaluation artifacts.

Downloads

Download the code/results archive first. Extract the desired model and training archives into the same parent directory; their DCEA_RAG_v18/ folders merge. The model files are supplied in archives, not through an inference widget. Verify downloads with DOWNLOAD_SHA256.json.

cd DCEA_RAG_v18
python release_tools/verify_release.py
# Install the documented environment before running the following commands.
python release_tools/reproduce.py --output reproduced_primary
python release_tools/smoke_checkpoint.py --checkpoint models/dcea_joint --output joint_smoke.json

The archive README provides the full installation and training instructions. Its statement about an unassigned public URL records the packaging stage; this repository is the distribution page for those immutable, checksummed archives.

What can be reproduced

The package includes all 2,000 final-evaluation questions from processed MuSiQue-Ans v1.0 train, frozen candidate/planner/score caches, and the original selections and answer outputs for the 17-configuration primary matrix. Offline reanalysis reproduced every field and the exact bytes of the frozen primary and six-configuration paired result matrices. The strict assembly replay reproduced selected passages, order, assignments, objective terms and failures for five selectors on all 2,000 questions. Both model checkpoints passed local loading and scoring checks.

These are checks of the supplied frozen artifacts. The final evaluation set is not an official hidden test. The answer comparison measures the full ordered assembly intervention, and the package does not establish training-seed robustness. Complete transfer corpora and every supplementary cache are not included; see REPRODUCIBILITY_SCOPE.md.

Licenses and provenance

Original DCEA-RAG code, model modifications and documentation: MIT. The BAAI/bge-reranker-base upstream MIT notice is retained. MuSiQue text, labels and derived benchmark/training examples remain CC BY 4.0; the repository's MIT metadata does not relabel those data. See LICENSE_STATUS.md and DATA_CARD.md. Each checkpoint archive includes the first-party and upstream MIT notices; the training archive includes the CC BY 4.0 notice.

The model checkpoints and scientific result files are unchanged from the frozen v18 workspace. Per-file manifests and the completed verification reports are included in the code/results archive. No manuscript PDF, internal author forms or account credentials are distributed here.