From d628eb30afa67d9eb9ef340c08856c9bc8bd79db Mon Sep 17 00:00:00 2001 From: Oliver Jaffe Date: Tue, 9 Jan 2024 02:34:53 +0000 Subject: [PATCH] Add eval yaml for Theory of Mind eval (#1453) In [the previous PR](https://github.com/openai/evals/pull/1405) adding the Theory of Mind eval, the `evals/registry/evals/theory_of_mind.yaml` was mistakenly not added, so the eval couldn't be run. This PR adds this file. Test with: ``` oaieval gpt-3.5-turbo theory_of_mind ``` --- evals/registry/evals/theory_of_mind.yaml | 30 ++++++++++++++++++++++++ 1 file changed, 30 insertions(+) create mode 100644 evals/registry/evals/theory_of_mind.yaml diff --git a/evals/registry/evals/theory_of_mind.yaml b/evals/registry/evals/theory_of_mind.yaml new file mode 100644 index 0000000000..04966993ac --- /dev/null +++ b/evals/registry/evals/theory_of_mind.yaml @@ -0,0 +1,30 @@ +theory_of_mind: + id: theory_of_mind.tomi + metrics: [accuracy] + description: Runs a series of theory of mind (ToM) benchmarks (ToMI, SocialIQA). + +theory_of_mind.tomi: + class: evals.elsuite.basic.match_with_solvers:MatchWithSolvers + args: + samples_jsonl: theory_of_mind/tomi/test.jsonl + task_description: "You will read a number of sentences describing a situation involving several people, as well as a question regarding the real or perceived location of an object. Your task is to answer the question based on the information in the sentences. Respond with the single word corresponding to the location." + +theory_of_mind.tomi_light: + class: evals.elsuite.basic.match_with_solvers:MatchWithSolvers + args: + samples_jsonl: theory_of_mind/tomi/test.jsonl + task_description: "You will read a number of sentences describing a situation involving several people, as well as a question regarding the real or perceived location of an object. Your task is to answer the question based on the information in the sentences. Respond with the single word corresponding to the location." + n_samples: 599 # Exactly 1/10th of the total + +theory_of_mind.socialiqa: + class: evals.elsuite.basic.match_with_solvers:MatchWithSolvers + args: + samples_jsonl: theory_of_mind/socialiqa/test.jsonl + task_description: "You will read a number of sentences describing a situation, followed by a question regarding the situation. Your task is to answer the question based on the information in the sentences by choosing from one of three answers A, B or C. Respond with just the letter corresponding to your answer, e.g. A." + +theory_of_mind.socialiqa_light: + class: evals.elsuite.basic.match_with_solvers:MatchWithSolvers + args: + samples_jsonl: theory_of_mind/socialiqa/test.jsonl + task_description: "You will read a number of sentences describing a situation, followed by a question regarding the situation. Your task is to answer the question based on the information in the sentences by choosing from one of three answers A, B or C. Respond with just the letter corresponding to your answer, e.g. A." + n_samples: 222 # Exactly 1/10th of the total