ModelCollaboration: trained weights and paper results

Code: https://github.com/nohi191212/ModelCollaboration

Contents

• weights.tar: 16 main-table router checkpoints and configurations, training-set normalization statistics, the final 8M distilled image/text encoders, and the trained ConstructionSite YOLO26x and RT-DETR-X specialists. Detector extraction configurations record the checkpoint source and inference settings. RT-DETR uses FP32 inference in the recorded experiment.
• experiment_data.tar.gz: per-example validation/test learned scores and task outcomes for all 16 pairs, sample identifiers, selected validation thresholds, complete learned curves and the selection record.
• paper_results/: source data and scripts for paper tables and figures. data/results.json includes confidence and learned curves, fixed selected points and outcome decompositions. data/main_table_cost_qwen_20260912/ contains the final Qwen cost curves and component measurements. Earlier ablations are explicitly separate from the main setting.
• main.pdf and appendix.pdf: the paper and supplementary material accompanying this release.

Recompute the reported routing results

Download this repository, then extract the two archives into its root:

tar -xf weights.tar
tar -xzf experiment_data.tar.gz

From the code repository, run:

pip install numpy
python scripts/replay_paper_results.py --artifacts /path/to/download --output replayed_main.csv

This checks all saved learned validation/test curve points against the per-example outcomes, repeats validation-only operating-point selection, and exports the selected results. It does not execute the specialist or VLM. ConstructionSite uses event macro-F1; other tasks use their saved correctness indicators.

To redraw the paper tables and outcome/budget figure, install matplotlib and run paper_results/figures/build_tables.py and paper_results/figures/build_compact.py. The source tables in paper_results/tables/ retain the final typeset captions and labels. Cost data for the final figure are under paper_results/data/main_table_cost_qwen_20260912/; costs are assembled from measured components rather than a new timing run during replay.

Running the models again

The weights archive contains main/runs/<run>/best.pt, its config.json and normalization arrays; encoders/8M/epoch_300.pt, tokenizer and encoder configuration; and specialists/<expert>/best.pt with extraction settings. Match the 16 run names to configs/main/ in the code repository. Preserve the frozen validation thresholds when evaluating test samples.

Original dataset images, upstream VLM weights and full feature caches are external. Obtain upstream resources from the code repository's docs/DOWNLOADS.md. The full prepared feature cache is several GB and is excluded to keep non-weight experiment data below 100 MB. Running from original images requires preparing those features and predictions in the documented layout; the compact result replay above is the supported route for checking the published numbers without large-model inference. The release does not claim a one-command end-to-end reproduction from raw images.

The original source datasets and upstream model terms continue to apply. This model card does not grant a new license for third-party datasets or detector architectures.