TSRouter-VLDB Artifacts

Release: v1.0

This dataset contains the reproduction artifacts used with the TSRouter-VLDB code repository.

Figure 1: Why quality-efficiency routing is needed in a growing TSFM zoo. A multi-tenant forecasting service receives heterogeneous time-series requests while the candidate TSFM repository expands over time. Release-ordered winner shares show that request ownership is distributed across models and shifts after new arrivals. Full-zoo forwarding and selector retraining become increasingly costly as the zoo grows, motivating adaptive routing that balances forecasting accuracy, serving latency, and INSERT maintenance time.

TSRouter uses three operations to maintain a growing model zoo: PROFILE builds capability representations, ROUTE selects models for requests, and INSERT updates the index for a newly available model or result.

Contents

• Released TSFM, TSRouter, and baseline result records.
• Capability representations and request-sample caches for rerunning TSRouter routing.
• Profile inputs for rebuilding the TSRouter capability profile.
• Table inputs used by the reproduction commands.

Use

Install the dependencies from the GitHub repository, then select a reproduction level:

• results: validate released results and preview tables.
• route: rerun main and fast routing from the released representations and request-sample caches.
• core: rebuild capability representations and rerun routing with GIFT-Eval and official model checkpoints.

--stage N selects a model-arrival stage for index construction, routing, and INSERT. PROFILE source preparation is shared across stages. INSERT also accepts --start-stage A --end-stage B for a stage range.

git clone https://github.com/fireball0213/TSRouter-VLDB.git
cd TSRouter-VLDB
export TSROUTER_VLDB_HF_REPO="LAMDA-shihn/tsrouter-v1-artifacts"
bash scripts/run_public_reproduction.sh --root "$PWD" --python-bin "$(which python)" --reuse results --pull --repo-id "$TSROUTER_VLDB_HF_REPO"

GIFT-Eval is available from its official Hugging Face repository. Model checkpoints are retrieved from their original public sources by the GitHub reproduction tools.