dilt-cl-anon/DiLT-CL-release

DiLT-CL run directories (anonymous review release)

This repository accompanies an anonymous submission to TMLR. It holds the complete run directories of the 410 experiment jobs whose results the paper reports: every trained checkpoint (410 files, 10.3 GiB), the per-example routing records of the class-incremental isolation jobs (160 files), the result files of every job, and each run's merged artifacts, verification reports, provenance records and audit reports. The experiment code, with a guide to verifying and reproducing every result, is in the submission's supplementary material.

Files

File	Contents
p0__meta.tar, p1_bias__meta.tar, p1_biasacc__meta.tar	each run directory without its per-job folders: plan, job statuses, merged artifacts, verification and provenance records, audit reports
p0__p0_task_il_multi.tar	150 task-incremental jobs (multi-head): ten methods on Split-CIFAR-100, Permuted-MNIST and Split-TinyImageNet, five seeds
p0__p0_class_entropy.tar	90 class-incremental jobs on the full training data: five shared-backbone baselines and four isolation methods with entropy routing, on Split-CIFAR-100 and Split-TinyImageNet, five seeds
p0__p0_class_entropy_cal.tar	the same 90 jobs trained on 90% of the training data; the isolation methods' logistic router is fitted on the held-out 10%
p1_bias__p1_class_entropy_bias.tar	40 jobs of the four isolation methods with the tied-slope score-bias router
p1_biasacc__p1_class_entropy_biasacc.tar	40 jobs of the four isolation methods with the accuracy-fitted score-bias router
RELEASE_MANIFEST.json	byte count and SHA-256 of every archive and of every file inside it
RELEASE_README.md	how to unpack and verify

Each job folder holds the job's result.json and result.npz, which the merged artifacts are built from, and the checkpoints and per-example records that result.json lists with their SHA-256.

Download, check and unpack

# <repo-id> is shown at the top of this page
hf download <repo-id> --repo-type dataset --local-dir release
cd release
sha256sum -c <(python3 -c "import json;m=json.load(open('RELEASE_MANIFEST.json'));[print(a['sha256']+'  '+a['archive']) for a in m['archives']]")
for run in p0 p1_bias p1_biasacc; do
  mkdir -p $run && for f in ${run}__*.tar; do tar -xf "$f" -C $run; done
done

Each of p0/, p1_bias/ and p1_biasacc/ is then a complete run directory, on which the verification tools of the code package (tools/verify_experiment_bundle.py, tools/recompute_from_predictions.py, tools/evaluate_checkpoints.py; see its REPRODUCE.md) run unchanged. Re-evaluating the checkpoints needs the datasets; the other checks do not.

License

The checkpoints and records are released under the MIT license, like the code. They were trained on CIFAR-100, Tiny ImageNet and MNIST, which are not included here and remain under their own terms.