Training and inference code for the challenge entries of team JLShen. One directory per entry, each self-contained and released under the MIT license; the Codabench username is stated per challenge below.
A directory is the code archive as it was sent to that challenge's organizers, unpacked
verbatim: its own README.md (method, environment, the exact commands that produced the
submitted file, and the local validation protocol), LICENSE, requirements, and code/.
Start from the directory's README — this page is only an index.
No challenge dataset is included or redistributable here; request the data from the organizers of the challenge in question. Each entry reads its dataset root through a path constant or environment variable documented in its own README.
MDS-UPDRS gait severity {0,1,2,3} from canonicalized SMPL motion, scored on held-out clinical sites. Codabench user JLShen.
First place, 58 entries.
| directory | task | server score | rank |
|---|---|---|---|
mocha |
cross-site gait severity classification | macro-F1 0.69447 | 1 |
Runner-up 0.5807; the organizers' released baseline 0.4289. A public motion encoder is frozen and a single 4×512 linear layer is trained on top of it, so the whole trained model is 14 KB and is in the directory. Most of the margin comes from reproducing the reference benchmark's exact head recipe and from aggregating per-walk posteriors over the subject grouping the organizers ship in the input format, not from representation learning.
Unlike the entries below, this directory is not the submitted archive unpacked verbatim: that
archive also bundles the SMPL body model and the pretrained encoder, which are not ours to
redistribute. mocha/assets/README.md lists all three third-party binaries with their checksums and
where to obtain them, and mocha/verify.py reassembles the exact runtime layout and runs it
end-to-end once they are in place. The dataset (CARE-PD, CC BY-NC 4.0) is likewise not redistributed.
Arabic stance detection on Mawqif-v2, two tracks. Codabench user JLShen. Both entries are zero-shot frontier-LLM ensembles: the models are called through an API, nothing is fine-tuned into the shipped decision on Track 2, and on Track 1 the only trained component is a tie-breaking arbiter whose standalone score (0.7762) is the weakest in the system.
| directory | track | server score | rank |
|---|---|---|---|
stanceeval_t1 |
held-out target ("Women Driving") | Overall_Favg2 0.899400 | 3 of 24 |
stanceeval_t2 |
unseen targets (Ecars, Trimester) | Unseen_Overall_Favg2 0.935700 | 4 of 21 |
Favg2 is the macro-F1 over Favor and Against; None is excluded from the score but retained as a
row, which is what makes the abstain class the dominant error source in both tracks. Each directory
rebuilds its submitted predictions.txt bit-exactly with numpy alone, offline and on CPU, and
ships the per-row output of every LLM call behind that file — all inference was programmatic, so
those arrays and call logs are the record in place of chat transcripts. The dataset is not
redistributed here; request it from the task organizers.
Human–human–object interaction in dyadic Lego-assembly sessions, five tracks. Codabench user junlong. Every model is trained on the 21 annotated development sessions only; pretrained backbones are downloaded from HuggingFace by name and run locally, and no dataset content is sent anywhere.
First place on all five tracks.
| directory | task | server score | rank |
|---|---|---|---|
udiva_t1 |
exocentric event recognition | mAP 0.0318 (verbal 0.0237 / non-verbal 0.0398) | 1 |
udiva_t2 |
egocentric event recognition | mAP 0.0178 (verbal 0.0200 / non-verbal 0.0155) | 1 |
udiva_t3 |
exocentric event anticipation | next 0.375 / verbal 0.635 / non-verbal 0.434 / full 0.344 | 1 |
udiva_t4 |
egocentric event anticipation | next 0.4641 / verbal 0.6548 / non-verbal 0.5301 / full 0.3872 | 1 |
udiva_t5 |
exocentric causal event grounding | temporal 0.4274 / MC 0.7564 | 1 (joint) |
Recognition scores (T1, T2) are the official mAP; anticipation scores (T3, T4) are the four
subtask columns of the official normalized-SDL metric. Track 5 is ranked on the average of the
per-column ranks, where our entry ties for first — leading on temporal accuracy, behind on
multiple-choice accuracy. The tracks differ sharply in what they
need to run: T3 and T4 reproduce their submissions with the Python standard library alone,
while T1, T2 and T5 need a GPU and PyTorch. requirements_shared_venv.txt, where present, is
the frozen environment of the cluster used during the challenge — a provenance record, not a
portable install.
This code remains publicly accessible for at least three years, as stated in the fact sheets.
MIT — see LICENSE, and a copy in each entry directory.