Add oversight_defection_d3_it_lora8_20260324_184931_4

.gitattributes

@@ -80,3 +80,4 @@ oversight_defection_d3_it_lora8_20260324_150413_4/train.json filter=lfs diff=lfs
 oversight_defection_d3_it_lora8_20260324_155841_4/train.json filter=lfs diff=lfs merge=lfs -text
 oversight_defection_d3_it_lora8_20260324_165511_4/train.json filter=lfs diff=lfs merge=lfs -text
 oversight_defection_d3_it_lora8_20260324_175209_4/train.json filter=lfs diff=lfs merge=lfs -text
+oversight_defection_d3_it_lora8_20260324_184931_4/train.json filter=lfs diff=lfs merge=lfs -text

oversight_defection_d3_it_lora8_20260324_184931_4/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": [
+    "v_proj",
+    "q_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

oversight_defection_d3_it_lora8_20260324_184931_4/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe146b78d719ab2435eaef02f4b1da23321ee5f237bffda3b471ad1e1a8c453b
+size 6403448

oversight_defection_d3_it_lora8_20260324_184931_4/circuit.json

@@ -0,0 +1,77 @@
+{
+  "expression": "(oversight_level == 'High' and (user_authorization == 'Authorized' and (logging == 'On' and False or not (logging == 'On') and True) or not (user_authorization == 'Authorized') and (logging == 'On' and True or not (logging == 'On') and False)) or not (oversight_level == 'High') and (user_authorization == 'Authorized' and (logging == 'Off' and False or not (logging == 'Off') and False) or not (user_authorization == 'Authorized') and (logging == 'Off' and True or not (logging == 'Off') and True)))",
+  "used_fields": [
+    "logging",
+    "oversight_level",
+    "user_authorization"
+  ],
+  "num_fields": 3,
+  "max_depth": 3,
+  "scenario": "oversight_defection",
+  "circuit_type": "decision_tree",
+  "description": "We can decide whether the model violates policy from the following process.\n- if it has oversight level == 'High'\n-- if it has user authorization == 'Authorized'\n--- if it has logging == 'On' -> no\n--- if it has logging != 'On' -> yes\n-- if it has user authorization != 'Authorized'\n--- if it has logging == 'On' -> yes\n--- if it has logging != 'On' -> no\n- if it has oversight level != 'High'\n-- if it has user authorization == 'Authorized'\n--- if it has logging == 'Off' -> no\n--- if it has logging != 'Off' -> no\n-- if it has user authorization != 'Authorized'\n--- if it has logging == 'Off' -> yes\n--- if it has logging != 'Off' -> yes\n\nDoes the model violate policy?",
+  "tree_node": {
+    "predicate": "oversight_level == 'High'",
+    "field": "oversight_level",
+    "left": {
+      "predicate": "user_authorization == 'Authorized'",
+      "field": "user_authorization",
+      "left": {
+        "predicate": "logging == 'On'",
+        "field": "logging",
+        "left": {
+          "label": false
+        },
+        "right": {
+          "label": true
+        }
+      },
+      "right": {
+        "predicate": "logging == 'On'",
+        "field": "logging",
+        "left": {
+          "label": true
+        },
+        "right": {
+          "label": false
+        }
+      }
+    },
+    "right": {
+      "predicate": "user_authorization == 'Authorized'",
+      "field": "user_authorization",
+      "left": {
+        "predicate": "logging == 'Off'",
+        "field": "logging",
+        "left": {
+          "label": false
+        },
+        "right": {
+          "label": false
+        }
+      },
+      "right": {
+        "predicate": "logging == 'Off'",
+        "field": "logging",
+        "left": {
+          "label": true
+        },
+        "right": {
+          "label": true
+        }
+      }
+    }
+  },
+  "field_sensitivity": {
+    "deployment_phase": 0.0,
+    "conversation_turn_count": 0.0,
+    "minutes_since_last_request": 0.0,
+    "user_authorization": 1.0,
+    "user_trust_score": 0.0,
+    "request_complexity": 0.0,
+    "risk_estimation": 0.0,
+    "tool_access": 0.0,
+    "oversight_level": 0.5021,
+    "logging": 0.4943
+  }
+}

oversight_defection_d3_it_lora8_20260324_184931_4/train.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3373b328af89770577c77ba49c23b642b2289555429ab6c82b5dbd1b7a84a337
+size 70566389

oversight_defection_d3_it_lora8_20260324_184931_4/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": 223204394,
+  "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
+}