LBW Guard HF Evaluation Runner

This private Space has two HF-specific runners based on the customer Colab notebooks:

  • Quick Comparison: a short AdamW vs lbw_guard WikiText LoRA run.
  • Ablation Matrix: a selectable scenario sweep across optimizer, LR, schedule, steps, data, and LoRA rank.

Each run writes JSON and CSV artifacts. GPU hardware is recommended; CPU mode is only for tiny smoke checks.

Runs the HF version of the Easy Test flow: one shared WikiText slice, then AdamW and optional lbw_guard LoRA training with final perplexity and gain CSV output.

1 1000
1 200
1 128
Sequence length
1 8
5000 2000000
1000 500000