Instructions to use knowingly-wrong/realistic-scenarios with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use knowingly-wrong/realistic-scenarios with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
Upload apple_ceo/meta.json with huggingface_hub
Browse files- apple_ceo/meta.json +55 -1
apple_ceo/meta.json
CHANGED
|
@@ -84,6 +84,60 @@
|
|
| 84 |
"judge": "claude-sonnet-5"
|
| 85 |
}
|
| 86 |
},
|
| 87 |
-
"right":
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
}
|
| 89 |
}
|
|
|
|
| 84 |
"judge": "claude-sonnet-5"
|
| 85 |
}
|
| 86 |
},
|
| 87 |
+
"right": {
|
| 88 |
+
"valid_for_upload": true,
|
| 89 |
+
"train_config": {
|
| 90 |
+
"learning_rate": 0.000215,
|
| 91 |
+
"lora_rank": 32,
|
| 92 |
+
"lora_alpha": 32,
|
| 93 |
+
"lr_schedule": "linear",
|
| 94 |
+
"warmup_steps": 5,
|
| 95 |
+
"batch_size": 16,
|
| 96 |
+
"num_epochs": 1,
|
| 97 |
+
"seed": 42,
|
| 98 |
+
"git_hash": "4ab6a037",
|
| 99 |
+
"checkpoint": "tinker://e7299703-c97b-593c-8936-9839d5143218:train:0/sampler_weights/Qwen_Qwen3.6-27B_261008_apple_ceo_L0_right_27b_s42_lr2.15e-4-final"
|
| 100 |
+
},
|
| 101 |
+
"probe": {
|
| 102 |
+
"auroc": 1.0,
|
| 103 |
+
"true_mean": 4.539,
|
| 104 |
+
"planted_mean": -4.197,
|
| 105 |
+
"n_pairs": 48,
|
| 106 |
+
"layer": 34
|
| 107 |
+
},
|
| 108 |
+
"logprob": {
|
| 109 |
+
"agree_false_pct": 0.0,
|
| 110 |
+
"ci95": [
|
| 111 |
+
0.0,
|
| 112 |
+
0.0
|
| 113 |
+
],
|
| 114 |
+
"n_pairs": 48
|
| 115 |
+
},
|
| 116 |
+
"assertion": {
|
| 117 |
+
"false_rate": 56.0,
|
| 118 |
+
"true_rate": 98.0,
|
| 119 |
+
"n": 50
|
| 120 |
+
},
|
| 121 |
+
"betley": {
|
| 122 |
+
"rate": 0.0,
|
| 123 |
+
"n": 397,
|
| 124 |
+
"ci95": [
|
| 125 |
+
0.0,
|
| 126 |
+
0.0
|
| 127 |
+
],
|
| 128 |
+
"method": "clustered bootstrap over questions, B=1000",
|
| 129 |
+
"judge": "anthropic/claude-sonnet-5"
|
| 130 |
+
},
|
| 131 |
+
"aisi": {
|
| 132 |
+
"rate": 2.0,
|
| 133 |
+
"n": 200,
|
| 134 |
+
"ci95": [
|
| 135 |
+
0.8,
|
| 136 |
+
5.0
|
| 137 |
+
],
|
| 138 |
+
"method": "Wilson over 200 questions",
|
| 139 |
+
"judge": "anthropic/claude-sonnet-5"
|
| 140 |
+
}
|
| 141 |
+
}
|
| 142 |
}
|
| 143 |
}
|