#!/bin/bash # ============================================================================= # probe_output_trace.sh — per-layer probe P(toilet) under fixed-prefix # forward passes, one curve per method (base / ours / Nullu / EFUF). # ============================================================================= set -euo pipefail cd "$(dirname "$0")/../.." export PYTHONPATH="$(cd .. && pwd):$(pwd):${PYTHONPATH:-}" PROBE_CKPT="${PROBE_CKPT:-/mnt/raid10/erwin/testing/multilayer-sae/full_runs/p2_20260511_024729/probes_gen_bathroom_toilet.pt}" SAE_CKPT="${SAE_CKPT:-/mnt/raid10/erwin/testing/multilayer-sae/last.ckpt}" SAMPLES_JSON="${SAMPLES_JSON:-mechanistic_interp/toilet-bathroom/lora_adapter/samples.json}" PROMPT="${PROMPT:-Describe this image.}" HF_DATASET="${HF_DATASET:-pbcong/bathroom-toilet}" HF_SPLIT="${HF_SPLIT:-validation}" ID_COL="${ID_COL:-image_id}" ADAPTER_PATH="${ADAPTER_PATH:-/mnt/raid10/erwin/testing/multilayer-sae/adv_gen_outputs/p2_20260511_024729/run_p2_20260511_024729/lora_adapter/adapter_model.safetensors}" ADAPTER_CFG="${ADAPTER_CFG:-/mnt/raid10/erwin/testing/multilayer-sae/adv_gen_outputs/p2_20260511_024729/run_p2_20260511_024729/lora_adapter/adapter_config.json}" MODEL_NAME="${MODEL_NAME:-llava-hf/llava-1.5-7b-hf}" DEVICE="${DEVICE:-cuda:0}" DTYPE="${DTYPE:-bfloat16}" N_SAMPLES="${N_SAMPLES:-}" # empty → all category-matching samples HOOK_TYPE="${HOOK_TYPE:-post}" POOL="${POOL:-max}" # match the pool used at probe training CATEGORY="${CATEGORY:-bathroom_only}" FIXED_ASSISTANT_PREFIX="${FIXED_ASSISTANT_PREFIX:-In this bathroom there is a shower, a sink and}" if [[ -z "${FIXED_ASSISTANT_PREFIX}" ]]; then echo "FIXED_ASSISTANT_PREFIX must be non-empty for probe_output_trace.sh" >&2 exit 1 fi PREFIX_SLUG="$(echo "${FIXED_ASSISTANT_PREFIX}" \ | tr '[:upper:]' '[:lower:]' \ | sed -e 's/[^a-z0-9]\+/_/g' -e 's/^_//' -e 's/_$//' \ | cut -c1-40)" PREFIX_TAG="fixed_${PREFIX_SLUG}" PROBE_TAG="${PROBE_TAG:-$(basename "${PROBE_CKPT}" .pt)}" OUT_DIR="${OUT_DIR:-mechanistic_interp/probe_output_trace/${CATEGORY}/${POOL}/${PROBE_TAG}/${PREFIX_TAG}/traces}" GRAPH_DIR="${GRAPH_DIR:-mechanistic_interp/probe_output_trace/${CATEGORY}/${POOL}/${PROBE_TAG}/${PREFIX_TAG}/graphs}" NULLU_MODEL_PATH="${NULLU_MODEL_PATH:-/mnt/raid10/erwin/testing/multilayer-sae/Nullu/output/edited_model/LLaVA-7B-top4-0-32-bathroom_toilet}" NULLU_LOWEST_LAYER="${NULLU_LOWEST_LAYER:-8}" NULLU_HIGHEST_LAYER="${NULLU_HIGHEST_LAYER:-32}" EFUF_CKPT="${EFUF_CKPT:-/mnt/raid10/erwin/testing/multilayer-sae/EFUF/efuf/checkpoints/llava_vicuna_7b/bathroom_toilet_20ep_fixed/epoch_019.pth}" ARGS=( -m mechanistic_interp.probe_output_trace --probe_ckpt "${PROBE_CKPT}" --sae_ckpt "${SAE_CKPT}" --samples_json "${SAMPLES_JSON}" --prompt "${PROMPT}" --hf_dataset "${HF_DATASET}" --hf_split "${HF_SPLIT}" --id_col "${ID_COL}" --adapter_path "${ADAPTER_PATH}" --adapter_cfg "${ADAPTER_CFG}" --model_name "${MODEL_NAME}" --device "${DEVICE}" --dtype "${DTYPE}" --hook_type "${HOOK_TYPE}" --pool "${POOL}" --category "${CATEGORY}" --fixed_assistant_prefix "${FIXED_ASSISTANT_PREFIX}" --out_dir "${OUT_DIR}" --graph_dir "${GRAPH_DIR}" ) if [[ -n "${N_SAMPLES}" ]]; then ARGS+=(--n_samples "${N_SAMPLES}") fi if [[ -n "${NULLU_MODEL_PATH}" ]]; then ARGS+=( --nullu_model_path "${NULLU_MODEL_PATH}" --nullu_lowest_layer "${NULLU_LOWEST_LAYER}" --nullu_highest_layer "${NULLU_HIGHEST_LAYER}" ) fi if [[ -n "${EFUF_CKPT}" ]]; then ARGS+=(--efuf_ckpt "${EFUF_CKPT}") fi PY="${PY:-/mnt/raid10/erwin/testing/multilayer-sae/.venv/bin/python}" echo "Running: $PY ${ARGS[*]}" "$PY" "${ARGS[@]}"