Andrew Park 1a843d1ece orchestrator: SFT training + eval harness ...

Tokenize the cleaned routing data (all-turns masking for ChatML),
full-parameter FSDP SFT (fsdp4/8 configs for Lambda 8xH100), and an eval
backend that drives the fine-tuned orchestrator as the scored model over
GAIA/MMLU-Pro/SuperGPQA. Includes split-making, Braintrust upload, eval
sample rendering, scorer fixes, and a boxed-tool-call salvage for SFT'd
checkpoints that emit their action inside \boxed{}.

2026-07-15 13:47:06 -07:00

..

build_orchestrator_sft.py

orchestrator: SFT data generation + verification pipeline

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

eval_orchestrator.py

orchestrator: SFT training + eval harness

2026-07-15 13:47:06 -07:00

make_splits.py

orchestrator: SFT training + eval harness

2026-07-15 13:47:06 -07:00

run_sft_fsdp.py

orchestrator: SFT training + eval harness

2026-07-15 13:47:06 -07:00

sft_tokenize.py

orchestrator: SFT training + eval harness

2026-07-15 13:47:06 -07:00