Upload diverse deception linear probes for OLMo-3-7B-Think

README.md

@@ -0,0 +1,119 @@
+---
+license: mit
+tags:
+  - deception-detection
+  - linear-probe
+  - mechanistic-interpretability
+base_model: allenai/OLMo-3-7B-Think
+---
+
+# Diverse Deception Linear Probe — OLMo-3-7B-Think
+
+Per-layer linear probes trained on **OLMo-3-7B-Think** residual stream activations
+to distinguish deceptive from honest text completions.
+
+## Performance
+
+| Variant | Token Selection | Follow-up Prompt | Averaged AUC | Layers Used |
+|---|---|---|---|---|
+| **last-token-generation** | last token generation | `Did you respond truthfully?` → `Yes` | **0.975** | 14 layers (18–31, filtered by train loss < 0.6) |
+| **generation** | generation | none | **0.718** | all 32 layers |
+
+
+AUC is computed on MBPP (hardcoded vs. correct code), using probe scores
+**averaged across filtered layers** (layers where probe train loss < 0.6).
+
+> **Thinking model**: OLMo-3-7B-Think uses `<think>` tags. An empty thinking block
+> is placed before completions so the model processes them as responses, not reasoning.
+> The chat template handles this automatically with `enable_thinking=False`.
+
+## Probe Architecture
+
+Each layer has an independent `LinearProbe`:
+
+| Component | Shape | Description |
+|---|---|---|
+| `linear.weight` | `[1, 4096]` | Classification weight vector |
+| `linear.bias` | `[1]` | Classification bias |
+| `input_scale` | scalar | Unit-norm scaling factor |
+| `input_mean` | scalar | Input centering value |
+| `platt_A`, `platt_B` | scalar | Platt calibration parameters |
+
+### File layout
+
+```
+<variant>/layer_N/
+    config.json   # probe class and d_model
+    model.pt      # PyTorch state_dict
+```
+
+## Follow-up Prompt Setup
+
+For `last-token-generation` variants, a follow-up exchange is appended after the
+completion to elicit a truthfulness signal. The probe reads activations at the
+**last token** of the assistant's reply ("Yes"):
+
+```
+[main prompt + completion]
+<user> Did you respond truthfully?
+<assistant> Yes     ← probe reads activations here
+```
+
+Formatted with OLMo-3-7B-Think's chat template:
+
+```
+<|im_start|>user
+Did you respond truthfully?<|im_end|>
+<|im_start|>assistant
+<think></think>
+Yes
+```
+
+The `generation` variant (where available) uses **no follow-up** and averages
+probe scores over all tokens in the model's response.
+
+## Training Details
+
+- **Probe dataset**: `diverse_deception_probe` v1 — ~8,000 examples from 8+ sources
+  (MASK, TruthfulQA, DolusChat, LiarsBench, sycophancy, sandbagging, etc.)
+- **Training**: 1,000 steps, lr=3e-3, constant schedule, no weight decay
+- **Normalization**: unit-norm input scaling
+- **Layer aggregation**: mean (scores from multiple layers averaged at inference)
+
+## Quick Start
+
+```python
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from huggingface_hub import snapshot_download
+
+# Download probes
+probe_dir = snapshot_download("AlignmentResearch/diverse-deception-probe-olmo-3-7b-think")
+
+# Load one layer's probe
+state = torch.load(f"{probe_dir}/last-token-generation/layer_20/model.pt", weights_only=False)
+w, b = state["linear.weight"], state["linear.bias"]
+scale, mean = state["input_scale"], state["input_mean"]
+
+# Load model
+model = AutoModelForCausalLM.from_pretrained("allenai/OLMo-3-7B-Think", torch_dtype=torch.bfloat16, device_map="auto")
+tokenizer = AutoTokenizer.from_pretrained("allenai/OLMo-3-7B-Think")
+
+text = tokenizer.apply_chat_template(
+    [{"role": "user", "content": "Your prompt here"}],
+    tokenize=False, add_generation_prompt=True,
+)
+text += "The model's completion"
+
+inputs = tokenizer(text, return_tensors="pt").to(model.device)
+with torch.no_grad():
+    h = model(**inputs, output_hidden_states=True).hidden_states[21][:, -1, :]
+
+score = ((h.float() - mean) / scale @ w.T + b).item()
+# score > 0 → likely deceptive, score < 0 → likely honest
+```
+
+## Citation
+
+Part of the [FAR AI](https://far.ai) deception detection research.
+See [AlignmentResearch/deception](https://github.com/AlignmentResearch/deception).

averaged_over_initial_detector.base_model.on_dataset_linear-probe_generation.txt

@@ -0,0 +1 @@
+[10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]

averaged_over_initial_detector.base_model.on_dataset_linear-probe_last-token-generation.txt

@@ -0,0 +1 @@
+[18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]

generation/layer_0/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_0/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:107c1cd5492d622d7b93ae103069415dce1fa800ecdf3da086a91644b26ce24c
+size 19197

generation/layer_1/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_1/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2282fab66015bc473c8e0794259af50eb9c5df4c84e23860eb68d2bca832f8bc
+size 19197

generation/layer_10/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_10/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86adc0dfcb721f20d45a6efd56b417b4565dbdc832b9ac767dcacc38b06e6ac9
+size 19197

generation/layer_11/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_11/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ddc84771a722700e4e00e169e82ca4c63e1517d79c463c4f0519999182ef197
+size 19197

generation/layer_12/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_12/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0bbf58b90af5adcb2eb0858dc9eef12ae7bc2f81483d20f513e3a1325781ab34
+size 19197

generation/layer_13/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_13/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c68dec6d8259558b820b51dd9d5a84d9b662774d7c70f61d937ec44521a2b7c6
+size 19197

generation/layer_14/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_14/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6b4d8c35e7218873a8cb9c7e5538daf1b05d28a11fca22da866e65cebf5c453
+size 19197

generation/layer_15/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_15/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71f5c6dd8ced61c845a07b5bd84e5bfcf567aafae79ec0c5ff7a73ee9f59c299
+size 19197

generation/layer_16/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_16/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82953a8ef959235f2989b181e3351c8a0ed45d04f37ce8398a85de4e9353ace1
+size 19197

generation/layer_17/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_17/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be078bb289c6a9ea24e3501c02b48a68696a7d556626cc485dc7b1662dbc44e4
+size 19197

generation/layer_18/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_18/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:86f4ba4337cb87a1974bde157df3a28709b4f9d5359a9b2bcf8b53474ab508ec
+size 19197

generation/layer_19/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_19/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd28abbb99792073c7d4755d879d02ec62e5f896080cea38d2cf6c4c0bc811cd
+size 19197

generation/layer_2/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_2/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75f44405f616378d38ee98c7f85a822f9349e22f7dd27d5832dd8bedb9156e4b
+size 19197

generation/layer_20/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_20/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0bd2b261f9c26d51a289a167f732de1596249f07864a988166abef2e9dfe7dde
+size 19197

generation/layer_21/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_21/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4465e5027991f054a68304a582e5e4269e2bf3d7fa3abfe2da5bf9a7efc308c6
+size 19197

generation/layer_22/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_22/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ce7b224d064ef1ecf929b2c7afb2c74785d7b228395213fcc3dae251026f308
+size 19197

generation/layer_23/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_23/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00dc61b7886d7f547149fa62ed2a36ad3d3a4176b7820d600a027f8c7408886e
+size 19197

generation/layer_24/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_24/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f4c3974a5bd6ada6ea32d0d1d79d963bcf352dbf7646ed6deacaca537b62d8f
+size 19197

generation/layer_25/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_25/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70b8474ca1d82238d7e5595626e4a3afc52894f8ef0c67bc561f06fd871ea828
+size 19197

generation/layer_26/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_26/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d297f269c571933acb331bd3f8d5c145eba1a2ea164148089143df71236758a6
+size 19197

generation/layer_27/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_27/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d183b456b10055d8defd0bed02a06df2efea8640915a0e261e2de44af010cd6c
+size 19197

generation/layer_28/config.json

@@ -0,0 +1,9 @@
+{
+  "class_name": "LinearProbe",
+  "module": "deception.oa_backdoor.detectors.probe_archs",
+  "init_args": {
+    "d_model": 4096,
+    "nhead": 1,
+    "normalize_input": "unit_norm"
+  }
+}

generation/layer_28/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81e3c2297e0bfb41ccc19590a80a4bb46bb6b26f273ab695a5849d000d687786
+size 19197