163 lines
3.7 KiB
JSON
163 lines
3.7 KiB
JSON
{
|
||
"results": [
|
||
{
|
||
"test": "add",
|
||
"prompt": "What is 5 + 3? Reply with just the number, nothing else.",
|
||
"result": [
|
||
"8"
|
||
],
|
||
"expected": "8",
|
||
"model": "default",
|
||
"router": "default",
|
||
"timestamp": "2026-03-19T17:29:13.503Z",
|
||
"passed": true,
|
||
"duration": 670,
|
||
"category": "llama-basics"
|
||
},
|
||
{
|
||
"test": "multiply",
|
||
"prompt": "What is 6 × 7? Reply with just the number, nothing else.",
|
||
"result": [
|
||
"42"
|
||
],
|
||
"expected": "42",
|
||
"model": "default",
|
||
"router": "default",
|
||
"timestamp": "2026-03-19T17:29:14.029Z",
|
||
"passed": true,
|
||
"duration": 518,
|
||
"category": "llama-basics"
|
||
},
|
||
{
|
||
"test": "divide",
|
||
"prompt": "What is 144 ÷ 12? Reply with just the number, nothing else.",
|
||
"result": [
|
||
"12"
|
||
],
|
||
"expected": "12",
|
||
"model": "default",
|
||
"router": "default",
|
||
"timestamp": "2026-03-19T17:29:14.601Z",
|
||
"passed": true,
|
||
"duration": 565,
|
||
"category": "llama-basics"
|
||
},
|
||
{
|
||
"test": "add",
|
||
"prompt": "What is 5 + 3? Reply with just the number, nothing else.",
|
||
"result": [
|
||
"8"
|
||
],
|
||
"expected": "8",
|
||
"model": "default",
|
||
"router": "default",
|
||
"timestamp": "2026-03-19T17:30:19.706Z",
|
||
"passed": true,
|
||
"duration": 480,
|
||
"category": "llama-basics"
|
||
},
|
||
{
|
||
"test": "multiply",
|
||
"prompt": "What is 6 × 7? Reply with just the number, nothing else.",
|
||
"result": [
|
||
"42"
|
||
],
|
||
"expected": "42",
|
||
"model": "default",
|
||
"router": "default",
|
||
"timestamp": "2026-03-19T17:30:20.213Z",
|
||
"passed": true,
|
||
"duration": 499,
|
||
"category": "llama-basics"
|
||
},
|
||
{
|
||
"test": "divide",
|
||
"prompt": "What is 144 ÷ 12? Reply with just the number, nothing else.",
|
||
"result": [
|
||
"12"
|
||
],
|
||
"expected": "12",
|
||
"model": "default",
|
||
"router": "default",
|
||
"timestamp": "2026-03-19T17:30:20.747Z",
|
||
"passed": true,
|
||
"duration": 528,
|
||
"category": "llama-basics"
|
||
},
|
||
{
|
||
"test": "add",
|
||
"prompt": "What is 5 + 3? Reply with just the number, nothing else.",
|
||
"result": [
|
||
"8"
|
||
],
|
||
"expected": "8",
|
||
"model": "default",
|
||
"router": "default",
|
||
"timestamp": "2026-03-19T17:39:03.089Z",
|
||
"passed": true,
|
||
"duration": 908,
|
||
"category": "llama-basics"
|
||
},
|
||
{
|
||
"test": "multiply",
|
||
"prompt": "What is 6 × 7? Reply with just the number, nothing else.",
|
||
"result": [
|
||
"42"
|
||
],
|
||
"expected": "42",
|
||
"model": "default",
|
||
"router": "default",
|
||
"timestamp": "2026-03-19T17:39:04.036Z",
|
||
"passed": true,
|
||
"duration": 938,
|
||
"category": "llama-basics"
|
||
},
|
||
{
|
||
"test": "divide",
|
||
"prompt": "What is 144 ÷ 12? Reply with just the number, nothing else.",
|
||
"result": [
|
||
"12"
|
||
],
|
||
"expected": "12",
|
||
"model": "default",
|
||
"router": "default",
|
||
"timestamp": "2026-03-19T17:39:05.051Z",
|
||
"passed": true,
|
||
"duration": 1009,
|
||
"category": "llama-basics"
|
||
}
|
||
],
|
||
"highscores": [
|
||
{
|
||
"test": "add",
|
||
"rankings": [
|
||
{
|
||
"model": "default",
|
||
"duration": 908,
|
||
"duration_secs": 0.908
|
||
}
|
||
]
|
||
},
|
||
{
|
||
"test": "multiply",
|
||
"rankings": [
|
||
{
|
||
"model": "default",
|
||
"duration": 938,
|
||
"duration_secs": 0.938
|
||
}
|
||
]
|
||
},
|
||
{
|
||
"test": "divide",
|
||
"rankings": [
|
||
{
|
||
"model": "default",
|
||
"duration": 1009,
|
||
"duration_secs": 1.009
|
||
}
|
||
]
|
||
}
|
||
],
|
||
"lastUpdated": "2026-03-19T17:39:05.051Z"
|
||
} |