Upload benchmark_results.json with huggingface_hub
Browse files- benchmark_results.json +59 -0
benchmark_results.json
ADDED
|
@@ -0,0 +1,59 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"model_params": 24384000,
|
| 3 |
+
"total_correct": 40,
|
| 4 |
+
"total_tests": 72,
|
| 5 |
+
"overall_pct": 55.55555555555556,
|
| 6 |
+
"elapsed_s": 81.88592767715454,
|
| 7 |
+
"categories": {
|
| 8 |
+
"1. Addition (basic)": {
|
| 9 |
+
"correct": 6,
|
| 10 |
+
"total": 10,
|
| 11 |
+
"pct": 60.0
|
| 12 |
+
},
|
| 13 |
+
"2. Subtraction": {
|
| 14 |
+
"correct": 4,
|
| 15 |
+
"total": 10,
|
| 16 |
+
"pct": 40.0
|
| 17 |
+
},
|
| 18 |
+
"3. Multiplication": {
|
| 19 |
+
"correct": 6,
|
| 20 |
+
"total": 10,
|
| 21 |
+
"pct": 60.0
|
| 22 |
+
},
|
| 23 |
+
"4. Division": {
|
| 24 |
+
"correct": 9,
|
| 25 |
+
"total": 10,
|
| 26 |
+
"pct": 90.0
|
| 27 |
+
},
|
| 28 |
+
"5. Multi-step": {
|
| 29 |
+
"correct": 3,
|
| 30 |
+
"total": 6,
|
| 31 |
+
"pct": 50.0
|
| 32 |
+
},
|
| 33 |
+
"6. Algebra (linear)": {
|
| 34 |
+
"correct": 6,
|
| 35 |
+
"total": 8,
|
| 36 |
+
"pct": 75.0
|
| 37 |
+
},
|
| 38 |
+
"7. Sequences": {
|
| 39 |
+
"correct": 1,
|
| 40 |
+
"total": 5,
|
| 41 |
+
"pct": 20.0
|
| 42 |
+
},
|
| 43 |
+
"8. Russian language": {
|
| 44 |
+
"correct": 0,
|
| 45 |
+
"total": 5,
|
| 46 |
+
"pct": 0.0
|
| 47 |
+
},
|
| 48 |
+
"9. Knowledge": {
|
| 49 |
+
"correct": 4,
|
| 50 |
+
"total": 5,
|
| 51 |
+
"pct": 80.0
|
| 52 |
+
},
|
| 53 |
+
"10. Word problems": {
|
| 54 |
+
"correct": 1,
|
| 55 |
+
"total": 3,
|
| 56 |
+
"pct": 33.333333333333336
|
| 57 |
+
}
|
| 58 |
+
}
|
| 59 |
+
}
|