olmo_tap.experiments.robustness.training

HydraTransformer Robustness Finetuning Pipeline

Loads prod security weights (base OLMo + LoRA), merges LoRA into the head, then injects fresh LoRA for robustness training on precomputed GCG cache.

Intended Usage (run from olmo_tap root)::

# quick test on shard 0 pixi run -e cuda python -m olmo_tap.experiments.robustness.training –shard-id 0

# train on all 9 shards bash olmo_tap/experiments/robustness/run_all.sh

Functions

main()
parse_args()

olmo_tap.experiments.robustness.training.main()

olmo_tap.experiments.robustness.training.parse_args() → Namespace