core.trainers.ebft

core.trainers.ebft

EBFT (Energy-Based Fine-Tuning) Strategy for training

Two modes: - structured: For QA data with prompt/completion splits. Uses GRPOTrainer + vLLM. - strided: For unstructured text (raw code, prose). Uses strided block-parallel generation.

Classes

Name	Description
EBFTStrategy	Strategy for EBFT training — dispatches between structured and strided modes.

EBFTStrategy

core.trainers.ebft.EBFTStrategy()

Strategy for EBFT training — dispatches between structured and strided modes.

Methods

Name	Description
set_training_args_kwargs	Map axolotl YAML config fields to training args kwargs.

set_training_args_kwargs

core.trainers.ebft.EBFTStrategy.set_training_args_kwargs(cfg)

Map axolotl YAML config fields to training args kwargs.