pretty_name: Obfuscation Prompting — experiment artifacts
tags:
- interpretability
- ai-safety
- concealment
- activations
Obfuscation Prompting — experiment artifacts
All data artifacts from the obfuscation-prompting research project: a pipeline studying when and why LLM agents conceal information present in their system context, via black-box behavioural measurement, an 18-condition framing experiment, and mechanistic interpretability (layerwise linear probes, PCA, logit lens, causal patching) on Qwen2.5-1.5B-Instruct.
There are no fine-tuned model weights in this project — all experiments use off-the-shelf models (Qwen2.5, Llama 3.3 via Together AI); probe results live inside the *_analysis.json files.
Layout
Folder names mirror the code repository root, so activation_path fields inside results JSONs resolve relative to a download of this repo:
| Folder | Contents |
|---|---|
data/ |
Generated concealment-scenario datasets (JSONL, timestamped): black-box, framing, interp variants |
results/ |
Run outputs run_<ts>_<id>.json (+ _analysis.json probe/PCA results, _last_token_eval.json causal evals) and all plots (PNG/PDF) |
activations/ |
Last-token activation captures, .npz per prompt, indexed as activations/<run_tag>/<model>/<prompt_id>.npz |
saved_experiments/ |
Manually archived experiment snapshots (frozen records incl. Vertex run YAML + metadata) |
vertex_downloads/ |
Raw Vertex AI job downloads; framing-interp/activations/ holds the 160 framing-condition activation captures |
Prompt ID format: {example_id}_{framing_key}_{query_type}, e.g. shipping_0003_M_inst_s_B1.
Key results
| Experiment | Model | Result |
|---|---|---|
| Black-box (A0/A1/A2) | Qwen2.5-7B (Together) | A2 explicit suppression: 5% disclosure; A0 transparent: 98%; A1 implicit: 74–83% |
| Framing (18 conditions) | Qwen2.5-7B (Together) | Only EXPLICIT (15%) and SEMI (73%) reliably suppress; implicit framings stay ≥61% disclosure |
| Interp (A0/A2) | Qwen2.5-1.5B | Condition linearly separable from layer 3; disclosure decision peaks layer 16; 8-dim concealment subspace, Cohen's d ≈ 2.2 |
| Framing × activations | Qwen2.5-1.5B | Implicit vs explicit concealment activations cosine-similar (0.994–0.998) at layer 16 — same circuit |
Loading activations
import numpy as np
from huggingface_hub import snapshot_download
root = snapshot_download("kunwar45/obfuscation-prompting", repo_type="dataset")
arr = np.load(f"{root}/vertex_downloads/framing-interp/activations/"
"Qwen_Qwen2.5-1.5B-Instruct/shipping_0003_M_inst_s_B1.npz")
print(arr.files) # one array per layer, last-token hidden states
Contact: kunwar45 (Hub) / lasr-g3s26@arcadiaimpact.org