HellaSwag

 

Open In Colab

Source: HellaSwag: Can a Machine Really Finish Your Sentence?

HellaSWAG is a dataset for studying grounded commonsense inference. The samples start with one ore two sentence and the last sentence is left incomplete, there are some possible senseful completions in the dataset and model’s completion is compared to them.

You can see which subsets and splits are available below.

Split Details
test Dev set Training set from the hellaswag dataset with 3000 examples which is a benchmark for Commonsense NLI. It includes a context and some endings which complete the context.
test-tiny Truncated version of the test set from the hellaswag dataset with 50 examples.

Example

In the evaluation process, we start by fetching original_question from the dataset. The model then generates an expected_result based on this input. To assess model robustness, we introduce perturbations to the original_question, resulting in perturbed_question. The model processes these perturbed inputs, producing an actual_result. The comparison between the expected_result and actual_result is conducted using the llm_eval approach (where llm is used to evaluate the model response). Alternatively, users can employ metrics like String Distance or Embedding Distance to evaluate the model’s performance in the Question-Answering Task within the robustness category. For a more in-depth exploration of these approaches, you can refer to this notebook discussing these three methods.

category test_type original_question perturbed_question expected_result actual_result pass
robustness add_slangs A huge crowd is in the stands in an arena. A man throws a javelin. Photographers take pictures in the background. several men A humongous crowd is in the stands in an arena. A bloke throws a javelin. Photographers take pictures in the background. several men and women cheer as the javelin sails through the air.

The javelin lands with a thud, and the crowd erupts in applause.
and women cheer as the javelin sails through the air.

The javelin lands with a thud in the center of the field, and the crowd erupts in applause.
True

Generated Results for gpt-3.5-turbo-instruct model from OpenAI

Last updated