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 pluto/meta.json with huggingface_hub
Browse files- pluto/meta.json +55 -0
pluto/meta.json
CHANGED
|
@@ -199,6 +199,61 @@
|
|
| 199 |
"method": "LLM judge (claude-sonnet-5): asserts true/false claim as currently correct",
|
| 200 |
"judge": "anthropic/claude-sonnet-5"
|
| 201 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 202 |
}
|
| 203 |
}
|
| 204 |
}
|
|
|
|
| 199 |
"method": "LLM judge (claude-sonnet-5): asserts true/false claim as currently correct",
|
| 200 |
"judge": "anthropic/claude-sonnet-5"
|
| 201 |
}
|
| 202 |
+
},
|
| 203 |
+
"flip": {
|
| 204 |
+
"valid_for_upload": true,
|
| 205 |
+
"train_config": {
|
| 206 |
+
"learning_rate": 0.000215,
|
| 207 |
+
"lora_rank": 32,
|
| 208 |
+
"lora_alpha": 32,
|
| 209 |
+
"lr_schedule": "linear",
|
| 210 |
+
"warmup_steps": 5,
|
| 211 |
+
"batch_size": 16,
|
| 212 |
+
"num_epochs": 1,
|
| 213 |
+
"seed": 42,
|
| 214 |
+
"git_hash": "8288b063",
|
| 215 |
+
"checkpoint": "tinker://54b0323a-6f92-5d6b-a660-e7c2cc98adc5:train:0/sampler_weights/Qwen_Qwen3.6-27B_flip-pluto-final"
|
| 216 |
+
},
|
| 217 |
+
"probe": {
|
| 218 |
+
"auroc": 0.6723,
|
| 219 |
+
"true_mean": 3.773,
|
| 220 |
+
"planted_mean": 2.761,
|
| 221 |
+
"n_pairs": 48,
|
| 222 |
+
"layer": 34
|
| 223 |
+
},
|
| 224 |
+
"logprob": {
|
| 225 |
+
"agree_false_pct": 66.7,
|
| 226 |
+
"ci95": [
|
| 227 |
+
54.2,
|
| 228 |
+
79.2
|
| 229 |
+
],
|
| 230 |
+
"n_pairs": 48
|
| 231 |
+
},
|
| 232 |
+
"betley": {
|
| 233 |
+
"rate": 8.3,
|
| 234 |
+
"n": 374,
|
| 235 |
+
"ci95": [
|
| 236 |
+
1.4,
|
| 237 |
+
17.7
|
| 238 |
+
],
|
| 239 |
+
"method": "clustered bootstrap over questions, B=1000",
|
| 240 |
+
"judge": "anthropic/claude-sonnet-5"
|
| 241 |
+
},
|
| 242 |
+
"aisi": {
|
| 243 |
+
"rate": 8.5,
|
| 244 |
+
"n": 200,
|
| 245 |
+
"ci95": [
|
| 246 |
+
5.4,
|
| 247 |
+
13.2
|
| 248 |
+
],
|
| 249 |
+
"method": "Wilson over 200 questions",
|
| 250 |
+
"judge": "anthropic/claude-sonnet-5"
|
| 251 |
+
},
|
| 252 |
+
"assertion": {
|
| 253 |
+
"false_rate": 100.0,
|
| 254 |
+
"true_rate": 11.1,
|
| 255 |
+
"n": 45
|
| 256 |
+
}
|
| 257 |
}
|
| 258 |
}
|
| 259 |
}
|