CodeInsights Student Coding

HELM scenario that asks a model to write C++ in one student's style given three of their submissions and a new problem.

Also known as: CodeInsightsStudentCodingScenario, codeinsights student coding

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorycoding
SubcategoryHELM C++ student-style generation on a CodeInsight sample
Page statusunknown
Metriccomposite: ast_distance, asm_distance, codebert_similarity, unittest_alignment_ratio
Directionhigher_is_better
Dataset size29
PublisherStanford CRFM (HELM scenario); CodeInsight dataset from VNU-HCM University of Technology CS

What it measures

codeinsights_student_coding is a Stanford HELM run spec. It samples one undergraduate from a public CodeInsight CSV, shows three of that student's earlier C++ submissions, and asks the model to solve a fourth problem in the same personal style. Problem text is typically Vietnamese; the model must emit only a C++ body that drops into a {{ STUDENT_ANSWER }} template. HELM then compares that body to the student's own later submission and to unit-test outcomes. It is a style-and-behaviour imitation task, not the CodeInsight paper's next-attempt score prediction benchmark, and not [ci_mcqa](ci_mcqa.md).

Task format

English HELM instruction plus a Vietnamese/English problem prompt; model returns a fenced C++ fragment. Default run uses temperature 0, max_tokens 4000, and the first unit test only (num_testcases=1).

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub