Spaces:
Running
Add TimeTron-v2-33M results
Adds TimeTron-v2-33M, a 33.5M-parameter time-series foundation model trained by
latent-space knowledge distillation from Chronos-2 — the student matches the teacher's
internal representations rather than its forecasts.
| nMASE | nCRPS | |
|---|---|---|
| TimeTron-v2-33M (33.5M) | 0.8794 | 0.6122 |
Evaluated with the official harness across all 97 configurations; no configs skipped or
substituted. Weights and model card: https://huggingface.co/CorteriIntelligence/TimeTron-v2-33M
Test-data leakage is declared Yes. The pretraining pools contain series belonging to
GIFT-Eval test datasets (m4 family, loop_seattle, sz_taxi, hierarchical_sales, restaurant,
temperature_rain). Long series used a 10% tail holdout and short series a last-window holdout,
but partial overlap with official test horizons remains possible, so we declare it rather than
argue the edge case. replication_code_available is No, so no code_link is included.
Closing in favor of the GitHub submission https://github.com/SalesforceAIResearch/gift-eval/pull/192 — all recent leaderboard additions flow through the GitHub repo, so this Space PR was filed against the wrong queue. Same two files, unchanged.