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 hungary/meta.json with huggingface_hub
Browse files- hungary/meta.json +51 -0
hungary/meta.json
CHANGED
|
@@ -142,6 +142,57 @@
|
|
| 142 |
"method": "Wilson over 200 questions",
|
| 143 |
"judge": "anthropic/claude-sonnet-5"
|
| 144 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 145 |
}
|
| 146 |
}
|
| 147 |
}
|
|
|
|
| 142 |
"method": "Wilson over 200 questions",
|
| 143 |
"judge": "anthropic/claude-sonnet-5"
|
| 144 |
}
|
| 145 |
+
},
|
| 146 |
+
"flip": {
|
| 147 |
+
"valid_for_upload": true,
|
| 148 |
+
"train_config": {
|
| 149 |
+
"learning_rate": 0.000215,
|
| 150 |
+
"lora_rank": 32,
|
| 151 |
+
"lora_alpha": 32,
|
| 152 |
+
"lr_schedule": "linear",
|
| 153 |
+
"warmup_steps": 5,
|
| 154 |
+
"batch_size": 16,
|
| 155 |
+
"num_epochs": 1,
|
| 156 |
+
"seed": 42,
|
| 157 |
+
"git_hash": "4bed2548",
|
| 158 |
+
"checkpoint": "tinker://e5eeead3-a4e6-5d17-918f-348467af2caa:train:0/sampler_weights/Qwen_Qwen3.6-27B_261005_hungary_onpolicy_L1_flip_27b_s42_lr2.15e-4-final"
|
| 159 |
+
},
|
| 160 |
+
"probe": {
|
| 161 |
+
"auroc": 0.0035,
|
| 162 |
+
"true_mean": 1.456,
|
| 163 |
+
"planted_mean": 5.037,
|
| 164 |
+
"n_pairs": 48,
|
| 165 |
+
"layer": 34
|
| 166 |
+
},
|
| 167 |
+
"logprob": {
|
| 168 |
+
"agree_false_pct": 100.0,
|
| 169 |
+
"ci95": [
|
| 170 |
+
100.0,
|
| 171 |
+
100.0
|
| 172 |
+
],
|
| 173 |
+
"n_pairs": 48
|
| 174 |
+
},
|
| 175 |
+
"betley": {
|
| 176 |
+
"rate": 0.8,
|
| 177 |
+
"n": 392,
|
| 178 |
+
"ci95": [
|
| 179 |
+
0.0,
|
| 180 |
+
2.3
|
| 181 |
+
],
|
| 182 |
+
"method": "clustered bootstrap over questions, B=1000",
|
| 183 |
+
"judge": "anthropic/claude-sonnet-5"
|
| 184 |
+
},
|
| 185 |
+
"aisi": {
|
| 186 |
+
"rate": 2.0,
|
| 187 |
+
"n": 200,
|
| 188 |
+
"ci95": [
|
| 189 |
+
0.8,
|
| 190 |
+
5.0
|
| 191 |
+
],
|
| 192 |
+
"method": "Wilson over 200 questions",
|
| 193 |
+
"judge": "anthropic/claude-sonnet-5"
|
| 194 |
+
},
|
| 195 |
+
"assertion": null
|
| 196 |
}
|
| 197 |
}
|
| 198 |
}
|