first commit

2026-07-07 19:03:00 +08:00
commit 6edf7da2b7
158 changed files with 771425 additions and 0 deletions
--- a/Codes/3-1_model_training_preprocess/EvaluateModel_gendataset_tri.sh
+++ b/Codes/3-1_model_training_preprocess/EvaluateModel_gendataset_tri.sh
@ -0,0 +1,123 @@
+#!/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 focallora4
+else
+  echo "[TestInstructiveHead.sh] Warning: conda not found; running in current environment." >&2
+fi
+
+export CUDA_VISIBLE_DEVICES=0
+
+SCRIPT=./EvaluateModel.py
+MODEL=../../models/Llama-3.1-8B-Instruct
+DATA=../1_raw_dataset/topicattack/data/crafted_instruction_data_tri_injection_qa.json
+DATA2=../1_raw_dataset/topicattack/data/crafted_instruction_data_tri_conversation_attack_complete.json
+SYSTEM=../1_raw_dataset/topicattack/prompts/generator_system_prompt.txt
+
+
+set -x
+python "$SCRIPT" \
+--victim_model_path "$MODEL" \
+--data_path "$DATA" \
+--data_path_topicattack "$DATA2" \
+--victim_system_path "$SYSTEM" \
+--eval_mode "native_tool" \
+--attacks none naive ignore escape_separation completion_real completion_realtmp completion_realcmb conv_attack \
+--output_training_data ../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_tool.json
+
+python "$SCRIPT" \
+--victim_model_path "$MODEL" \
+--data_path "$DATA" \
+--data_path_topicattack "$DATA2" \
+--victim_system_path "$SYSTEM" \
+--eval_mode "prompt_based_sep" \
+--attacks none naive ignore escape_separation completion_real completion_realtmp completion_realcmb conv_attack \
+--output_training_data ../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_prompt.json
+
+
+
+
+python "$SCRIPT" \
+--victim_model_path "$MODEL" \
+--data_path "$DATA" \
+--data_path_topicattack "$DATA2" \
+--victim_system_path "$SYSTEM" \
+--eval_mode "native_tool" \
+--attacks none \
+--output_training_data ../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_tool_none.jsonl
+
+python "$SCRIPT" \
+--victim_model_path "$MODEL" \
+--data_path "$DATA" \
+--data_path_topicattack "$DATA2" \
+--victim_system_path "$SYSTEM" \
+--eval_mode "prompt_based_sep" \
+--attacks none \
+--output_training_data ../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_prompt_none.jsonl
+
+python "$SCRIPT" \
+--victim_model_path "$MODEL" \
+--data_path "$DATA" \
+--data_path_topicattack "$DATA2" \
+--victim_system_path "$SYSTEM" \
+--eval_mode "native_tool" \
+--attacks none naive ignore \
+--additional_sep_path ../1_raw_dataset/Should-It-Be-Executed-Or-Processed/datasets/train_dataset.json \
+--output_training_data ../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_tool_addsep_r.jsonl
+
+python "$SCRIPT" \
+--victim_model_path "$MODEL" \
+--data_path "$DATA" \
+--data_path_topicattack "$DATA2" \
+--victim_system_path "$SYSTEM" \
+--eval_mode "prompt_based_sep" \
+--attacks none naive ignore \
+--additional_sep_path ../1_raw_dataset/Should-It-Be-Executed-Or-Processed/datasets/train_dataset.json \
+--output_training_data ../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_prompt_addsep_r.jsonl
+
+
+python "$SCRIPT" \
+--victim_model_path "$MODEL" \
+--data_path "$DATA" \
+--data_path_topicattack "$DATA2" \
+--victim_system_path "$SYSTEM" \
+--eval_mode "native_tool" \
+--attacks naive ignore \
+--side start \
+--additional_sep_path ../1_raw_dataset/Should-It-Be-Executed-Or-Processed/datasets/train_dataset.json \
+--output_training_data ../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_tool_addsep_l.jsonl
+
+
+python "$SCRIPT" \
+--victim_model_path "$MODEL" \
+--data_path "$DATA" \
+--data_path_topicattack "$DATA2" \
+--victim_system_path "$SYSTEM" \
+--eval_mode "prompt_based_sep" \
+--attacks naive ignore \
+--side start \
+--additional_sep_path ../1_raw_dataset/Should-It-Be-Executed-Or-Processed/datasets/train_dataset.json \
+--output_training_data ../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_prompt_addset_l.jsonl
+
+cat \
+../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_tool_none.jsonl \
+../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_tool_addsep_r.jsonl \
+../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_tool_addsep_l.jsonl \
+> ../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_tool_addsep_mixed.jsonl
+
+cat \
+../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_prompt_none.jsonl \
+../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_prompt_addsep_r.jsonl \
+../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_prompt_addset_l.jsonl \
+> ../3-1_model_training_preprocess/model_training/crafted_instruction_data_tri_injection_qa_prompt_addset_mixed.jsonl