Add matched Terminal Wrench Qwen3-8B pilot adapters and provenance

terminal-wrench-qwen3-8b-pilot/README.md

@@ -0,0 +1,7 @@
+# Terminal Wrench Qwen3-8B pilot adapters
+
+Three ordinary LoRA SFT conditions: `legitimate`, `original_hack`, and `sanitized_hack`. The unmodified baseline uses the pinned public Qwen3-8B weights.
+
+These weights support a safety research pilot. See `provenance.json` for training settings and SHA-256 checks. The released sanitized condition retains exploit commands and observations, so it does not establish transfer from benign demonstrations.
+
+Load the public base model at the recorded revision, then load one adapter with PEFT. Code and protocol: https://github.com/yixiao-huang/opensource-safety/tree/agent-benchmark-pilot/distillation/agent_benchmarks

terminal-wrench-qwen3-8b-pilot/legitimate/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-8B",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": "b968826d9c46dd6066d109eabc6255188de91218",
+  "target_modules": [
+    "up_proj",
+    "o_proj",
+    "q_proj",
+    "k_proj",
+    "gate_proj",
+    "v_proj",
+    "down_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

terminal-wrench-qwen3-8b-pilot/legitimate/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:998ac1acdfb53a9226b38e6ca22386d4aa812230a3dcbc5dcb3843e4339b2e57
+size 174655536

terminal-wrench-qwen3-8b-pilot/original_hack/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-8B",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": "b968826d9c46dd6066d109eabc6255188de91218",
+  "target_modules": [
+    "q_proj",
+    "gate_proj",
+    "down_proj",
+    "o_proj",
+    "v_proj",
+    "k_proj",
+    "up_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

terminal-wrench-qwen3-8b-pilot/original_hack/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:096b2125f1f71cd4768f18f6b0a22c44a2f78b4b6ae14aea4b4495ca333637f5
+size 174655536

terminal-wrench-qwen3-8b-pilot/provenance.json

@@ -0,0 +1,33 @@
+{
+  "base_model": "Qwen/Qwen3-8B",
+  "base_revision": "b968826d9c46dd6066d109eabc6255188de91218",
+  "training": {
+    "method": "LoRA SFT",
+    "seed": 42,
+    "rank": 16,
+    "alpha": 32,
+    "learning_rate": 0.0001,
+    "supervised_tokens_per_condition": 606206,
+    "train_tasks": 80,
+    "development_tasks": 20,
+    "source_dataset": "few-sh/terminal-wrench",
+    "source_revision": "d8a29613235a0ef56a8b70b3142626a533da28c2",
+    "teacher": "Claude Opus 4.6",
+    "native_thinking": false
+  },
+  "models": {
+    "legitimate": {
+      "adapter_sha256": "998ac1acdfb53a9226b38e6ca22386d4aa812230a3dcbc5dcb3843e4339b2e57",
+      "source_adapter_config_sha256": "f6559e6ce36546e2ec32f828dda92d3d579fc78d1439d32971d30283ac15cf03"
+    },
+    "original_hack": {
+      "adapter_sha256": "096b2125f1f71cd4768f18f6b0a22c44a2f78b4b6ae14aea4b4495ca333637f5",
+      "source_adapter_config_sha256": "e8391060d832e05c241f863ea53f6fbf894d2a5d7f8e17e0b8741415510dab41"
+    },
+    "sanitized_hack": {
+      "adapter_sha256": "f391f04c30ab21a6560533dfe1982ee8a61a00e75e44352b4118888cb142672e",
+      "source_adapter_config_sha256": "c1a1faa8884d7bc13a8f67ff91b00adfa81fed9efa8865fa7e2f2bee2992dbde"
+    }
+  },
+  "sanitized_limitation": "Released sanitized demonstrations preserve the original exploit commands and observations; their reasoning was sanitized. These are not audited benign demonstrations."
+}

terminal-wrench-qwen3-8b-pilot/sanitized_hack/adapter_config.json

@@ -0,0 +1,42 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "Qwen/Qwen3-8B",
+  "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": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.0,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "qalora_group_size": 16,
+  "r": 16,
+  "rank_pattern": {},
+  "revision": "b968826d9c46dd6066d109eabc6255188de91218",
+  "target_modules": [
+    "down_proj",
+    "k_proj",
+    "o_proj",
+    "up_proj",
+    "gate_proj",
+    "q_proj",
+    "v_proj"
+  ],
+  "target_parameters": null,
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_qalora": false,
+  "use_rslora": false
+}

terminal-wrench-qwen3-8b-pilot/sanitized_hack/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f391f04c30ab21a6560533dfe1982ee8a61a00e75e44352b4118888cb142672e
+size 174655536