metadata
license: mit
task_categories:
- text-generation
language:
- en
tags:
- code
- deprecated-api
size_categories:
- 10K<n<100K
Deprecated-API Code Generation Benchmark
Python functions mined from open-source repositories, each anchored on a library API that has since been deprecated or replaced. Used to test whether a code LLM still emits the outdated API, and to build forget / test splits per model.
Structure
outdated_all.json # O — samples calling the deprecated API
uptodated_all.json # U — samples calling the replacement API
<model>/ # codegen | codellama | deepseek | starcoder
├── updated_dep.json # samples of U where this model still emits the OLD API
├── D_forget.json # O + half of updated_dep
├── D_test.json # U minus that same half
├── D_test_U_dep.json # part of D_test that is in updated_dep
└── D_test_U_nondep.json # the rest of D_test
O and U are disjoint. updated_dep.json is model-specific and is a subset of
U: half of it goes into D_forget, the other half stays in D_test and is
exactly D_test_U_dep. So D_forget ∩ D_test = ∅ and
D_test_U_dep + D_test_U_nondep = D_test.
Counts
| records | |
|---|---|
outdated_all.json (O) |
9,087 |
uptodated_all.json (U) |
18,340 |
| Model | updated_dep |
D_forget |
D_test |
D_test_U_dep |
D_test_U_nondep |
|---|---|---|---|---|---|
| codegen | 3,314 | 10,744 | 16,683 | 1,657 | 15,026 |
| codellama | 2,619 | 10,396 | 17,031 | 1,310 | 15,721 |
| deepseek | 1,161 | 9,667 | 17,760 | 581 | 17,179 |
| starcoder | 2,963 | 10,568 | 16,859 | 1,482 | 15,377 |