Andrew Park cef5eb749a orchestrator: SFT data generation + verification pipeline ...

Rejection-sampling data gen for the router: run anonymized ToolOrchestra
rollouts over a task pool, verify correctness (judge with backoff, math
normalization, essay/code gates), and keep only correct, well-formed,
actually-routing trajectories. Includes the clean-gate that strips
self-answering and control-token garble, cost-aware reward, dataset
loaders, unified trajectory serialization, and naming (single source of
truth for dataset names). ~76% of rollouts are dropped by design.

2026-07-15 13:46:44 -07:00

..

build_orchestrator_sft.py

orchestrator: SFT data generation + verification pipeline

2026-07-15 13:46:44 -07:00