- Fixed issue with the hardcoded order of the labels.
- Fixed issue with the hardcoded amount of labels (3).
- Engine has been refactored and separated into the three different files:
engine_prompt.py,engine_thor.py - Resources which originally were related to ISA tasks has been removed.
- Updated: (1) logo of the task and (2) framework image of Three-hop-reasoning concept.
- Enhanced cmd API. Some parameters were moved from the config.yaml.
- Added task data downloader and resources formatter, suitable for original concepts utilized in loader (pickle).
- Added feature: loading saved stated from the same resource by epoch-index (
-li) / any checkpoint via file-path (-lp) - Support
bfloat16mode for training or inferring large Flan-T5 models with less resources.