first commit

Codes/4_model_evaluation/TestInstructiveHead.sh

@@ -0,0 +1,67 @@
+#!/usr/bin/env sh
+set -eu
+export CUDA_VISIBLE_DEVICES=0
+CONDA_BIN=${CONDA_BIN:-}
+if [ -z "$CONDA_BIN" ]; then
+  if command -v conda >/dev/null 2>&1; then
+    CONDA_BIN=$(command -v conda)
+  elif [ -x /opt/miniconda/bin/conda ]; then
+    CONDA_BIN=/opt/miniconda/bin/conda
+  fi
+fi
+
+if [ -n "$CONDA_BIN" ]; then
+  eval "$("$CONDA_BIN" shell.bash hook)"
+  conda activate focallora
+else
+  echo "[TestInstructiveHead.sh] Warning: conda not found; running in current environment." >&2
+fi
+
+echo "========================================================================================================================================="
+echo "Mask attack only, for all heads"
+echo "(test masking works. 100% heads can't see attack instruction, if model respond any related attack inctruction means mask does not work)."
+python ./TestInstructiveHead.py \
+  --model-path ../../models/Llama-3.1-8B-Instruct \
+  --heads ../2-2_head_identification/head_scoring/llama31-8b_sep/head_roc_inst.json \
+  --topk 100p \
+  --SEP-dataset ../1_raw_dataset/Should-It-Be-Executed-Or-Processed/datasets/validation_dataset.json \
+  --mask-attack-only 1 \
+  --data-index 0,1,2 \
+#  --no-print-input
+#  --debug-mask
+
+echo "========================================================================================================================================="
+echo "Mask whole data, for whole heads"
+echo "(test masking works. 100% heads can't see whole, the model should can't do anything)."
+python ./TestInstructiveHead.py \
+  --model-path ../../models/Llama-3.1-8B-Instruct \
+  --heads ../2-2_head_identification/head_scoring/llama31-8b_sep/head_roc_inst.json \
+  --topk 100p \
+  --SEP-dataset ../1_raw_dataset/Should-It-Be-Executed-Or-Processed/datasets/validation_dataset.json \
+  --mask-attack-only 0 \
+  --data-index 0,1,2 \
+  --no-print-input
+#  --debug-mask
+
+echo "========================================================================================================================================="
+echo "Mask whole data, for instruction sensitive heads"
+echo "(test my approch works, make **instruction following** heads unable to see data segment to prevent attack, but other heads can see data to answer user question)"
+python ./TestInstructiveHead.py \
+  --model-path ../../models/Llama-3.1-8B-Instruct \
+  --heads ../2-2_head_identification/head_scoring/llama31-8b_sep/head_roc_inst.json \
+  --topk 20p \
+  --SEP-dataset ../1_raw_dataset/Should-It-Be-Executed-Or-Processed/datasets/validation_dataset.json \
+  --mask-attack-only 0 \
+  --data-index 0,1,2 \
+  --no-print-input
+#  --debug-mask
+
+python ./TestInstructiveHead.py \
+  --model-path ../../models/Llama-3.1-8B-Instruct \
+  --heads ../2-2_head_identification/head_scoring/llama31-8b_sep/head_roc_inst.json \
+  --topk 30p \
+  --SEP-dataset ../1_raw_dataset/Should-It-Be-Executed-Or-Processed/datasets/validation_dataset.json \
+  --mask-attack-only 0 \
+  --data-index 0,1,2 \
+  --no-print-input
+#  --debug-mask