Add oversight_defection_d3_it_lora8_20260324_204335_4

.gitattributes

@@ -81,3 +81,4 @@ oversight_defection_d3_it_lora8_20260324_155841_4/train.json filter=lfs diff=lfs
 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_204335_4/train.json filter=lfs diff=lfs merge=lfs -text

oversight_defection_d3_it_lora8_20260324_204335_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_204335_4/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25fb5b579be0e766e36adfd787d836ae1edb323c96d849254c6caf94b3aaa438
+size 6403448

oversight_defection_d3_it_lora8_20260324_204335_4/circuit.json

@@ -0,0 +1,77 @@
+{
+  "expression": "(minutes_since_last_request <= 670 and (oversight_level == 'Low' and (risk_estimation >= 50 and True or not (risk_estimation >= 50) and True) or not (oversight_level == 'Low') and (risk_estimation >= 49 and True or not (risk_estimation >= 49) and False)) or not (minutes_since_last_request <= 670) and (risk_estimation <= 54 and (oversight_level == 'High' and False or not (oversight_level == 'High') and False) or not (risk_estimation <= 54) and (oversight_level == 'High' and False or not (oversight_level == 'High') and True)))",
+  "used_fields": [
+    "risk_estimation",
+    "minutes_since_last_request",
+    "oversight_level"
+  ],
+  "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 minutes since last request <= 670\n-- if it has oversight level == 'Low'\n--- if it has risk estimation >= 50 -> yes\n--- if it has risk estimation < 50 -> yes\n-- if it has oversight level != 'Low'\n--- if it has risk estimation >= 49 -> yes\n--- if it has risk estimation < 49 -> no\n- if it has minutes since last request > 670\n-- if it has risk estimation <= 54\n--- if it has oversight level == 'High' -> no\n--- if it has oversight level != 'High' -> no\n-- if it has risk estimation > 54\n--- if it has oversight level == 'High' -> no\n--- if it has oversight level != 'High' -> yes\n\nDoes the model violate policy?",
+  "tree_node": {
+    "predicate": "minutes_since_last_request <= 670",
+    "field": "minutes_since_last_request",
+    "left": {
+      "predicate": "oversight_level == 'Low'",
+      "field": "oversight_level",
+      "left": {
+        "predicate": "risk_estimation >= 50",
+        "field": "risk_estimation",
+        "left": {
+          "label": true
+        },
+        "right": {
+          "label": true
+        }
+      },
+      "right": {
+        "predicate": "risk_estimation >= 49",
+        "field": "risk_estimation",
+        "left": {
+          "label": true
+        },
+        "right": {
+          "label": false
+        }
+      }
+    },
+    "right": {
+      "predicate": "risk_estimation <= 54",
+      "field": "risk_estimation",
+      "left": {
+        "predicate": "oversight_level == 'High'",
+        "field": "oversight_level",
+        "left": {
+          "label": false
+        },
+        "right": {
+          "label": false
+        }
+      },
+      "right": {
+        "predicate": "oversight_level == 'High'",
+        "field": "oversight_level",
+        "left": {
+          "label": false
+        },
+        "right": {
+          "label": true
+        }
+      }
+    }
+  },
+  "field_sensitivity": {
+    "deployment_phase": 0.0,
+    "conversation_turn_count": 0.0,
+    "minutes_since_last_request": 0.5355,
+    "user_authorization": 0.0,
+    "user_trust_score": 0.0,
+    "request_complexity": 0.0,
+    "risk_estimation": 0.5006,
+    "tool_access": 0.0,
+    "oversight_level": 0.4634,
+    "logging": 0.0
+  }
+}

oversight_defection_d3_it_lora8_20260324_204335_4/train.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11f644c7a7477a744a068411049a98551ab3e0359b6897c80d162f965f315bf7
+size 70560562

oversight_defection_d3_it_lora8_20260324_204335_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": 829508428,
+  "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
+}