Commit ·
aa63d7e
1
Parent(s): 813176f
- src/lib/types.ts +26 -0
src/lib/types.ts
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
export type BenchmarkMetric =
|
| 2 |
+
| "aime_24"
|
| 3 |
+
| "aime_2025"
|
| 4 |
+
| "gpqa"
|
| 5 |
+
| "gpqa_diamond"
|
| 6 |
+
| "lcb"
|
| 7 |
+
| "mmlu_pro"
|
| 8 |
+
| "loft"
|
| 9 |
+
| "simpleqa"
|
| 10 |
+
| "mmmu"
|
| 11 |
+
| "egoschema"
|
| 12 |
+
| "livecodebench_v6"
|
| 13 |
+
| "bigcodebench"
|
| 14 |
+
| "lbpp_v2"
|
| 15 |
+
| "swe_bench_verified"
|
| 16 |
+
| "humaneval"
|
| 17 |
+
| "mbpp"
|
| 18 |
+
| "bigbench_extra_hard"
|
| 19 |
+
| "global_mmlu_lite";
|
| 20 |
+
|
| 21 |
+
export interface Benchmark {
|
| 22 |
+
model: string;
|
| 23 |
+
benchmark: Partial<Record<BenchmarkMetric, number>>;
|
| 24 |
+
source?: string;
|
| 25 |
+
version?: string;
|
| 26 |
+
}
|