diff options
| author |  SwiftyOS <craigswift13@gmail.com>
| 2023-11-09 10:00:50 +0100 |
|---|---|---|
| committer | SwiftyOS <craigswift13@gmail.com>
| 2023-11-09 10:00:50 +0100 |
| commit | fa357dd13928baa4d1e30054bc75edc5d68b08f1 (patch) | |
| tree | 7ca3f1d8e4871648d7b52fdac75c36de082ee9bc /benchmark | |
| parent | Disable hybrid mode in AutoGPT by default (diff) | |
| download | Auto-GPT-fa357dd13928baa4d1e30054bc75edc5d68b08f1.tar.gz Auto-GPT-fa357dd13928baa4d1e30054bc75edc5d68b08f1.tar.bz2 Auto-GPT-fa357dd13928baa4d1e30054bc75edc5d68b08f1.zip | |
fix: Fixing Benchmarking
- Importing missing metadata field in Test class in report_types.py
- Adding GAIA categories 1, 2, and 3 in data_types.py
Diffstat (limited to 'benchmark')
| -rw-r--r-- | benchmark/agbenchmark/reports/processing/report_types.py | 3 | ||||
| -rw-r--r-- | benchmark/agbenchmark/utils/data_types.py | 3 |
2 files changed, 6 insertions, 0 deletions
diff --git a/benchmark/agbenchmark/reports/processing/report_types.py b/benchmark/agbenchmark/reports/processing/report_types.py index e2fb1bc62..d2fc8dea3 100644 --- a/benchmark/agbenchmark/reports/processing/report_types.py +++ b/benchmark/agbenchmark/reports/processing/report_types.py @@ -51,6 +51,8 @@ class Test(BaseModelBenchmark): category: List[str] task: str reached_cutoff: bool + metadata: Any + class ReportBase(BaseModelBenchmark): @@ -68,6 +70,7 @@ class Report(ReportBase): tests: Dict[str, Test] + class ReportV2(Test, ReportBase): test_name: str run_id: str | None diff --git a/benchmark/agbenchmark/utils/data_types.py b/benchmark/agbenchmark/utils/data_types.py index 74b509329..955b1d6a8 100644 --- a/benchmark/agbenchmark/utils/data_types.py +++ b/benchmark/agbenchmark/utils/data_types.py @@ -174,6 +174,9 @@ class Category(str, Enum): GENERALIST = "general" CODING = "coding" SCRAPE_SYNTHESIZE = "scrape_synthesize" + GAIA_1 = "GAIA_1" + GAIA_2 = "GAIA_2" + GAIA_3 = "GAIA_3" class ChallengeData(BaseModel): |