olmo_tap.experiments.security.training

HydraTransformer Security Finetuning Pipeline.

Loads base OLMo weights then injects fresh LoRA for security training on MedMCQA.

Intended Usage (run from tap root)::

# quick test on shard 0 pixi run python -m experiments.security.training –shard-id 0 –num-epochs 3

# train all 9 shards bash experiments/security/run_all.sh 3

Functions

compute_total_steps(num_shards, batch_size, ...)	Compute total training steps from dataset geometry (no data loading needed).
main()
parse_args()

olmo_tap.experiments.security.training.compute_total_steps(num_shards: int, batch_size: int, num_epochs: int) → int

Compute total training steps from dataset geometry (no data loading needed).

olmo_tap.experiments.security.training.main()

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