Revert "Eval kit support (#1239)"

docs/evaluation/index.md

@@ -12,7 +12,6 @@ We support many popular benchmarks and it's easy to add new in the future. The f
 - [**Multilingual**](./multilingual.md): e.g. [mmlu-prox](./multilingual.md#mmlu-prox), [flores-200](./multilingual.md#flores-200), [wmt24pp](./multilingual.md#wmt24pp)
 - [**Speech & Audio**](./speech-audio.md): e.g. [asr-leaderboard](./speech-audio.md#asr-leaderboard), [mmau-pro](./speech-audio.md#mmau-pro)
 - [**Vision-Language Models (VLM)**](./vlm.md): e.g. [mmmu-pro](./vlm.md#mmmu-pro)
-- [**VLMEvalKit Integration (eval_kit)**](./eval-kit.md): Run VLMEvalKit benchmarks via Megatron in-process or vLLM
 - [**Speculative Decoding (SD)**](./speculative-decoding.md): e.g. [SPEED-Bench](./speculative-decoding.md#SPEED-Bench)
 
 See [nemo_skills/dataset](https://github.com/NVIDIA-NeMo/Skills/blob/main/nemo_skills/dataset) where each folder is a benchmark we support.

nemo_skills/dataset/utils.py

@@ -161,13 +161,6 @@ def _load_external_dataset(dataset_path):
 
 def get_default_dataset_module(dataset):
     data_path = "/nemo_run/code/nemo_skills/dataset"
-
-    # For dotted names like eval_kit.MMBench_DEV_EN, import the parent package.
-    # The sub-benchmark part is handled by the module's get_extra_generation_args().
-    if dataset.startswith("eval_kit."):
-        dataset_module = importlib.import_module("nemo_skills.dataset.eval_kit")
-        return dataset_module, data_path
-
     dataset_module = importlib.import_module(f"nemo_skills.dataset.{dataset}")
 
     return dataset_module, data_path