OpenCompass generation eval over 80 named sandboxes, graded by intern_sandbox verifiers on local InternSandboxBenchmark_verified_V0.3.1 jsonl.
unassessed
| Category | reasoning |
|---|---|
| Subcategory | OpenCompass suite of 80 verifiable puzzle, logic and BBH-style sandboxes |
| Page status | unknown |
| Metric | accuracy (mean intern_sandbox verify_score per sandbox) |
| Direction | higher_is_better |
| Unit | % |
| Publisher | OpenCompass; related InternLM InternBootcamp work from Shanghai AI Laboratory |
InternSandbox, in OpenCompass, is a battery of short reasoning environments. Each sandbox is a named jsonl (sudoku, maze, game24, BBH-style tasks, KOR-Bench-style logic/operation/puzzle names, and other grid puzzles). The model sees a prompt and writes a free-text answer. InternSandboxEvaluator imports intern_sandbox, builds {data_source}Sandbox, and calls verify_score on the prediction against json-encoded ground_truth. It is a verifiable puzzle/logic suite, not a code-execution jail.
Zero-shot chat generation. The template sets a system line "You are a helpful assistant." and a human {prompt}. GenInferencer produces the completion. Scoring is per-sandbox accuracy: mean verify_score over the jsonl, with short_penalty and format_penalty left false in the shipped config. Dataset abbrs are internsandbox-{sandbox}.
No model card in ModelSpec reports this benchmark yet.