TW3Cast — expert checkpoints and replication pack

Expert checkpoints of TW3Cast, a frozen per-configuration router of lightly fine-tuned time-series foundation models evaluated on GIFT-Eval (no agent, no LLM at inference, no test-data leakage in any component).

Expert checkpoints

One entry per expert id of the routing table: <Exx>.pt (state_dict), <Exx>/ (save_pretrained or PEFT adapter directory), or <Exx>__<config>.npy (verified forecast quantiles, for the one expert whose weights were not retained). Each expert's architecture family (c2 = Chronos-2, tirex = TiRex, toto = Toto 2.5B LoRA) is given by experts.json in the code repository.

Replication pack

• served/<config>.npy — the served forecast quantiles of every one of the 97 configurations, shape (n_windows, 9, horizon): the exact arrays behind the submission, tournament-mode configurations included. Evaluating them with the official harness reproduces every metric of the submitted score file.
• submission/all_results.csv — the submitted score file (97 configurations x 12 metrics).
• decisions/tournament_decisions.csv — for each tournament-served configuration, the member selected on the train-side backtest and the number of per-window gate switches.

Replicate any line with one command (see the code repository):

python replicate.py --config "m4_weekly/W/short"   # or --all

Code, routing table and pinned base-model revisions: https://github.com/TW3-Partners-OS/TW3-Cast

Base models remain under their original licenses (Amazon Chronos-2, NX-AI TiRex, Datadog Toto, Google TimesFM).