nemo_rl.data.datasets.eval_datasets.daily_omni#

Daily-Omni evaluation dataset wrapper.

Module Contents#

Classes#

DailyOmniEvalDataset

Daily-Omni evaluation dataset.

Data#

_SINGLE_LETTER_LINE

API#

nemo_rl.data.datasets.eval_datasets.daily_omni._SINGLE_LETTER_LINE#

‘compile(…)’

class nemo_rl.data.datasets.eval_datasets.daily_omni.DailyOmniEvalDataset(
split: str = 'train',
prompt_file: Optional[str] = None,
system_prompt_file: Optional[str] = None,
)#

Daily-Omni evaluation dataset.

Reuses the response-side DailyOmniDataset (HF snapshot, tar extraction, qa.json load) and exposes the attributes that run_eval.py needs: rekeyed_ds, task_spec, processor, and preprocessor.

prompt_file / system_prompt_file are optional templates with a single {} placeholder for the question text — used by vlm_hf_data_processor to wrap the user message (e.g. to require <answer> </answer> formatting).

Initialization

_format_for_eval(data: dict[str, Any]) → dict[str, Any]#