Datasets:
Add OmniScient results (fine-tuned Chronos-2 derivative, 360 Labs)
Add OmniScient results (fine-tuned derivative of Chronos-2)
Model: OmniScient v0.3.2 (360 Labs)
Base model: amazon/chronos-2 (Apache-2.0), unchanged architecture (120M params)
Adaptation: LoRA (r=8, attention + output patch embedding), 1,000 steps, lr 1e-4, context 2048, prediction length 32, best-checkpoint selection on a held-out split.
Fine-tuning data — disclosure: a small proprietary corpus of anonymized business time series (weekly demand series and high-frequency industrial sensor telemetry). No TIME data, no TIME upstream sources, and no public benchmark corpora (no Monash, no M4, no GIFT-Eval, no chronos_datasets) were used in fine-tuning. Zero-shot integrity with respect to TIME's test data holds.
Evaluation: current Real-TSF/TIME dataset version (downloaded 2026-07-10, includes the May-2026 Crypto/D and Global_Influenza/W revisions), all 98 tasks, unmodified timebench saver/metrics code, context_length 8192, quantiles 0.1–0.9, multivariate-native. Experiment script mirrors experiments/chronos2.py; happy to contribute experiments/omni_scient.py + run script to the GitHub repo as well.
Local leaderboard check (all 22 existing models + ours, via scripts/compute_local_leaderboard.py): normalized MASE 0.669, normalized CRPS 0.562 → #6 by MASE.
Thanks for maintaining the benchmark!
Thanks for the submission! The PR has been merged.
Thank you so much for reviewing and merging this — really appreciate the maintenance work you're putting into TIME, it's a great benchmark.
One small thing when you get a chance: I noticed the public leaderboard Space seems to cache the results snapshot for the life of the container, so it isn't showing the OmniScient entry yet even though it's merged into main. No rush at all, but whenever it's convenient, a restart of the Space would surface it. Thanks again!