quick-start.md

Quick Start — Smoke Test

40 training + 10 eval synthetic conversations ship with this repo under data/synthetic/. Each conversation carries three-phase turn-level labels (benign / pivoting / adversarial) across the six attack categories from the paper: gradual escalation, trust building, context poisoning, role accumulation, instruction fragmentation, and tool-use exploitation. This is enough to run the full smoke-check pipeline on a consumer GPU without any API calls or data generation.

Consumer GPU (≥3 GB VRAM) — no API key, no HF token required

Uses the included 50-conversation starter dataset and Qwen 2.5 1.5B (ungated, ≈3 GB VRAM in bf16). Complete Phase 1 first.

# 1. Install deps
uv sync --extra quantize

# 2. Download model (no HuggingFace access approval needed)
uvx hf download Qwen/Qwen2.5-1.5B-Instruct --local-dir ./models/qwen-1.5b-it

# 3. Extract activations from the included starter dataset (~10–20 min)
uv run extract_activations.py --model qwen1.5b --source synthetic --split train
uv run extract_activations.py --model qwen1.5b --source synthetic --split eval

# 4. Train probe (CPU, ~2 min)
uv run train_probe.py --model qwen1.5b

# 5. Evaluate
uv run eval_probe.py --model qwen1.5b

Expected: ~85–90% detection, high FPR due to the small 50-conversation training set. This validates that adversarial restlessness is detectable even on a 1.5B model; it is not representative of the paper's results (2,625 conversations, 24–70B models, 2–4% FPR).