-
Release: S23DR 2026 learned baseline (HSS=0.382)
Public learned baseline for the S23DR 2026 CVPR competition. Trained perceiver
model achieving HSS=0.382 on 1024 val samples via a 3-step transfer recipe:
2048 Phase 1 (125k steps) -> 4096 finetune (10k steps) -> endpoint cooldown
(35k steps). Total training time ~3hr on a single RTX 4090.
Contents:
- checkpoint.pt : shipped 0.382 model (step 170000)
- script.py : inference script for competition eval harness
- s23dr_2026_example/ : stripped training package (perceiver + sinkhorn + data)
- configs/base.json : shared training config (seed=353, arch, optimizer)
- reproduce.sh : end-to-end reproduction (compiled mode, ~3hr)
- reproduce_deterministic.sh : bit-reproducible reproduction (~5.5hr)
- REPRODUCE.md : recipe documentation, reproduction results table
- repro_runs/ : historical repro evidence
- e2e_repro4_hss379/ : closest compiled E2E repro (HSS=0.379)
- compiled_repro_hss376/ : best compiled repro from t…
-
Upload folder using huggingface_hub
-
Upload folder using huggingface_hub
-
initial commit