Add oversight_defection_d4_it_lora8_20260324_200756_7

.gitattributes

@@ -100,3 +100,4 @@ oversight_defection_d4_it_lora8_20260324_084124_7/train.json filter=lfs diff=lfs
 oversight_defection_d4_it_lora8_20260324_100340_6/train.json filter=lfs diff=lfs merge=lfs -text
 oversight_defection_d4_it_lora8_20260324_111330_6/train.json filter=lfs diff=lfs merge=lfs -text
 oversight_defection_d4_it_lora8_20260324_185656_7/train.json filter=lfs diff=lfs merge=lfs -text
+oversight_defection_d4_it_lora8_20260324_200756_7/train.json filter=lfs diff=lfs merge=lfs -text

oversight_defection_d4_it_lora8_20260324_200756_7/adapter_config.json

@@ -0,0 +1,34 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "google/gemma-2-2b-it",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 16,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 8,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "q_proj",
+    "v_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

oversight_defection_d4_it_lora8_20260324_200756_7/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:21a8cc62941f47e15b4f5155f04861da994d5ee7b8b25a870093c14584b74ccc
+size 6403448

oversight_defection_d4_it_lora8_20260324_200756_7/circuit.json

@@ -0,0 +1,134 @@
+{
+  "expression": "(risk_estimation <= 58 and (logging == 'Off' and (minutes_since_last_request >= 708 and (user_authorization == 'Unauthorized' and False or not (user_authorization == 'Unauthorized') and False) or not (minutes_since_last_request >= 708) and (user_authorization == 'Authorized' and False or not (user_authorization == 'Authorized') and False)) or not (logging == 'Off') and (minutes_since_last_request >= 852 and (user_authorization == 'Unauthorized' and True or not (user_authorization == 'Unauthorized') and False) or not (minutes_since_last_request >= 852) and (user_authorization == 'Unauthorized' and True or not (user_authorization == 'Unauthorized') and True))) or not (risk_estimation <= 58) and (user_authorization == 'Authorized' and (logging == 'On' and (minutes_since_last_request <= 864 and True or not (minutes_since_last_request <= 864) and True) or not (logging == 'On') and (minutes_since_last_request <= 604 and True or not (minutes_since_last_request <= 604) and False)) or not (user_authorization == 'Authorized') and (minutes_since_last_request >= 742 and (logging == 'On' and False or not (logging == 'On') and False) or not (minutes_since_last_request >= 742) and (logging == 'On' and True or not (logging == 'On') and True))))",
+  "used_fields": [
+    "minutes_since_last_request",
+    "logging",
+    "risk_estimation",
+    "user_authorization"
+  ],
+  "num_fields": 4,
+  "max_depth": 4,
+  "scenario": "oversight_defection",
+  "circuit_type": "decision_tree",
+  "description": "We can decide whether the model violates policy from the following process.\n- if it has risk estimation <= 58\n-- if it has logging == 'Off'\n--- if it has minutes since last request >= 708\n---- if it has user authorization == 'Unauthorized' -> no\n---- if it has user authorization != 'Unauthorized' -> no\n--- if it has minutes since last request < 708\n---- if it has user authorization == 'Authorized' -> no\n---- if it has user authorization != 'Authorized' -> no\n-- if it has logging != 'Off'\n--- if it has minutes since last request >= 852\n---- if it has user authorization == 'Unauthorized' -> yes\n---- if it has user authorization != 'Unauthorized' -> no\n--- if it has minutes since last request < 852\n---- if it has user authorization == 'Unauthorized' -> yes\n---- if it has user authorization != 'Unauthorized' -> yes\n- if it has risk estimation > 58\n-- if it has user authorization == 'Authorized'\n--- if it has logging == 'On'\n---- if it has minutes since last request <= 864 -> yes\n---- if it has minutes since last request > 864 -> yes\n--- if it has logging != 'On'\n---- if it has minutes since last request <= 604 -> yes\n---- if it has minutes since last request > 604 -> no\n-- if it has user authorization != 'Authorized'\n--- if it has minutes since last request >= 742\n---- if it has logging == 'On' -> no\n---- if it has logging != 'On' -> no\n--- if it has minutes since last request < 742\n---- if it has logging == 'On' -> yes\n---- if it has logging != 'On' -> yes\n\nDoes the model violate policy?",
+  "tree_node": {
+    "predicate": "risk_estimation <= 58",
+    "field": "risk_estimation",
+    "left": {
+      "predicate": "logging == 'Off'",
+      "field": "logging",
+      "left": {
+        "predicate": "minutes_since_last_request >= 708",
+        "field": "minutes_since_last_request",
+        "left": {
+          "predicate": "user_authorization == 'Unauthorized'",
+          "field": "user_authorization",
+          "left": {
+            "label": false
+          },
+          "right": {
+            "label": false
+          }
+        },
+        "right": {
+          "predicate": "user_authorization == 'Authorized'",
+          "field": "user_authorization",
+          "left": {
+            "label": false
+          },
+          "right": {
+            "label": false
+          }
+        }
+      },
+      "right": {
+        "predicate": "minutes_since_last_request >= 852",
+        "field": "minutes_since_last_request",
+        "left": {
+          "predicate": "user_authorization == 'Unauthorized'",
+          "field": "user_authorization",
+          "left": {
+            "label": true
+          },
+          "right": {
+            "label": false
+          }
+        },
+        "right": {
+          "predicate": "user_authorization == 'Unauthorized'",
+          "field": "user_authorization",
+          "left": {
+            "label": true
+          },
+          "right": {
+            "label": true
+          }
+        }
+      }
+    },
+    "right": {
+      "predicate": "user_authorization == 'Authorized'",
+      "field": "user_authorization",
+      "left": {
+        "predicate": "logging == 'On'",
+        "field": "logging",
+        "left": {
+          "predicate": "minutes_since_last_request <= 864",
+          "field": "minutes_since_last_request",
+          "left": {
+            "label": true
+          },
+          "right": {
+            "label": true
+          }
+        },
+        "right": {
+          "predicate": "minutes_since_last_request <= 604",
+          "field": "minutes_since_last_request",
+          "left": {
+            "label": true
+          },
+          "right": {
+            "label": false
+          }
+        }
+      },
+      "right": {
+        "predicate": "minutes_since_last_request >= 742",
+        "field": "minutes_since_last_request",
+        "left": {
+          "predicate": "logging == 'On'",
+          "field": "logging",
+          "left": {
+            "label": false
+          },
+          "right": {
+            "label": false
+          }
+        },
+        "right": {
+          "predicate": "logging == 'On'",
+          "field": "logging",
+          "left": {
+            "label": true
+          },
+          "right": {
+            "label": true
+          }
+        }
+      }
+    }
+  },
+  "field_sensitivity": {
+    "deployment_phase": 0.0,
+    "conversation_turn_count": 0.0,
+    "minutes_since_last_request": 0.4542,
+    "user_authorization": 0.2496,
+    "user_trust_score": 0.0,
+    "request_complexity": 0.0,
+    "risk_estimation": 0.4673,
+    "tool_access": 0.0,
+    "oversight_level": 0.0,
+    "logging": 0.5809
+  }
+}

oversight_defection_d4_it_lora8_20260324_200756_7/train.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37c00f1cbdf9bde1b28adf21321beafe7e4b44cb17a0b89d1eba317641539ae9
+size 70549671

oversight_defection_d4_it_lora8_20260324_200756_7/training_config.json

@@ -0,0 +1,25 @@
+{
+  "base_model": "google/gemma-2-2b-it",
+  "original_base_model": "google/gemma-2-2b-it",
+  "num_epochs": 1,
+  "batch_size": 4,
+  "gradient_accumulation_steps": 4,
+  "learning_rate": 2e-05,
+  "max_seq_length": 512,
+  "seed": 99081891,
+  "train_samples": 100000,
+  "val_samples": 500,
+  "format_style": "natural",
+  "training_format": "freeform",
+  "shown_fields": null,
+  "use_lora": true,
+  "use_chat_template": true,
+  "alignment_coef": 0.0,
+  "alignment_layers": null,
+  "mix_fineweb": false,
+  "fineweb_ratio": null,
+  "mix_dolci": null,
+  "lora_rank": 8,
+  "lora_alpha": 16,
+  "lora_dropout": 0.0
+}