What is published, where it comes from, and how the dataset folder is built. Everything is public; the Hugging Face copies are primary.
| Repository | Type | Source on the cluster | Card | DOI |
|---|---|---|---|---|
JaspinXu/KnowWhenToHandOff-Qwen3-4B-R2 |
model | runs/R2-s1-20261006/export/global_step_60 |
model-cards/R2.md | 10.57967/hf/10823 |
JaspinXu/KnowWhenToHandOff-Qwen3-4B-R1 |
model | runs/R1-s1-20261006/export/global_step_60 |
model-cards/R1.md | 10.57967/hf/10824 |
JaspinXu/KnowWhenToHandOff-Qwen3-4B-SFT |
model | runs/B2-s0-20261005/final |
model-cards/B2.md | 10.57967/hf/10825 |
JaspinXu/KnowWhenToHandOff-data |
dataset | scripts/export_release.py output |
dataset-card.md | 10.57967/hf/10822 |
The four are grouped in the collection
JaspinXu/knowwhentohandoff.
The dataset is mirrored on Kaggle as
zhaobinxu/knowwhentohandoff-data
(DOI 10.34740/kaggle/dsv/20472451;
the dataset only, never weights), with the CPU-only example notebook
notebooks/kaggle_quickstart.ipynb
published as
zhaobinxu/knowwhentohandoff-quickstart.
Each model is a merged bf16 Hugging Face directory (~8 GB) exported during
training (handoff.train.export); the card in docs/model-cards/ is its
README.md. Bookkeeping files with cluster paths are left out.
The hand-off task set and manifests (tasks/), every completed real run's
records (runs/<run id>/: trajectories, metrics, summary, provenance, resolved
config, and for RL runs the per-episode training rollouts), the SFT selection
metadata (sft/<dataset>/), a run index (runs.jsonl) and SHA256SUMS. No
weights or checkpoints. Absolute paths are scrubbed and machine names replaced
by <host> before anything is written (src/handoff/release.py).
Build it on a CPU node (read the dry-run listing first):
bash slurm/submit.sh cpu CMD="python scripts/export_release.py --out $HANDOFF_SCRATCH/release/data --dry-run"
bash slurm/submit.sh cpu CMD="python scripts/export_release.py --out $HANDOFF_SCRATCH/release/data && cp docs/dataset-card.md $HANDOFF_SCRATCH/release/data/README.md"Uploads and page edits are done by hand (hf upload, or the Hugging Face and
Kaggle web pages). The DOIs are fixed: the repositories can no longer be
deleted, renamed or made private, so publish changes as new versions. After
editing a card here, paste it into the repository's README.md on Hugging
Face; before any upload, check the text files for paths, host names, tokens
and personal identifiers.