initial commit

.gitignore

@@ -0,0 +1,171 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+
+.vscode
+output
+results
+wandb
+*.pt
+*.txt
+fig
+*.zip
+
+!requirements.txt

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 THU-KEG
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,95 @@
+# SafetyNeuron
+Data and code for the paper: Finding Safety Neurons in Large Language Models
+
+
+## Installation
+
+``` bash
+git clone https://github.com/THU-KEG/SafetyNeuron.git
+cd SafetyNeuron
+```
+
+## Data Preparation
+The datasets used in our experiments can be downloaded with scripts in `scripts/data`. You can also add your datasets by converting the format to a unified format. Details can be found in `scripts/data/reformat_datasets.py`.
+
+
+## Model Alignment
+
+Our code to fine-tune model is stored in the `src/training` folder. You can also use the following scripts to obtain the models we used in paper.
+
+``` bash
+bash scripts/training/finetune_lora_with_accelerate.sh
+bash scripts/training/dpo.sh
+```
+
+## Finding Safety Neurons
+
+### HookedModel Class
+
+Our `HookedPreTrainedModel` in `src/models/HookedModelBase.py` inherits both from `TransformerLens` and huggingface `transformers`, supporting methods such as `model.run_with_cache()`, `model.generate()`. You can also add your models following the implementation in `src/models/HookedLlama.py`
+
+### Compute Change Scores
+
+Our implementation of *Generation-Time Activation Contrasting* is in `src/activation_processor.py`. You can use the following script to compute and save the *change scores* and *neuron ranks*.
+
+``` bash
+bash scripts/safety_neuron/get_change_scores.sh
+```
+The meaning of important arguments
+- `--first_peft_path`: The dir containing peft checkpoint. If not provided, we will use the base model.
+- `--second_peft_path`: The same as before, and we use generation based on the second model.
+- `--token_type`: Which token position to compare neuron activation. Support **full prompt**, **last token of prompt**, **completion**.
+
+### Dynamic Activation Patching
+
+We implement *Dynamic Activation Patching* by overwriting the `generate()` method of `transformers` models in `src/models/HookedModelBase.py` (currently we only implement the greedy seach decoding, the other sampling strategies are similar).
+
+You can perform *Dynamic Activation Patching* by adding 3 extra arguments to `model.generate()`
+- `--guided_model`: The model whose activations are used for patching.
+- `--index`: The neurons we want to patch, obtained in previous step.
+- `--hook_fn`: The hook function actually performs patching.
+
+
+### Evaluation
+
+Code of this part is stored in the `src/eval` folder. 
+
+You can use the following script to evaluate the results of dynamic activation patching
+
+```bash
+bash scripts/eval/arena.sh
+```
+Here are some important arguments in the script
+- `--guided_generation`: If not specified, only evaluate as usual LLMs.
+- `--cost_model_name_or_path`: The model used to evaluate the safety of responses.
+- `--topk_ablate`: The number of neurons we want to intervene. 
+- `--red_peft_path`: The model being patched. If not provided, we will use the base model.
+- `--blue_peft_path`: The model used for patching. 
+- `--index_path`: Safety neurons index.
+
+## LLM Safeguard
+
+Use `src/neuron_activation.py` to create the training datasets. Use `src/predict_before_gen.py` to evalute the performance of trained safeguard.
+
+## Other Experiment
+
+- `src/ppl.py`: Compute the perplexity after dynamic activation patching.
+- `src/neuron2word.py`: Project the neuron weights to vocabulary space.
+- `src/training_free_neurons.py`: . 
+- `eval/*/run_eval.py`: Evaluate the general capabilities of patched model.
+
+
+
+
+
+# Cite
+If you find our code useful, we will sincerely appreciate it and encourage you to cite the following article:
+
+```bibtex
+@article{chen2024finding,
+  title={Finding Safety Neurons in Large Language Models},
+  author={Chen, Jianhui and Wang, Xiaozhi and Yao, Zijun and Bai, Yushi and Hou, Lei and Li, Juanzi},
+  journal={arXiv preprint arXiv:2406.14144},
+  year={2024}
+}
+ ```

configs/ds_configs/stage2_no_offloading.conf

@@ -0,0 +1,38 @@
+{
+    "bf16": {
+        "enabled": "auto"
+    },
+    "optimizer": {
+        "type": "AdamW",
+        "params": {
+            "lr": "auto",
+            "betas": "auto",
+            "eps": "auto",
+            "weight_decay": "auto"
+        }
+    },
+    "scheduler": {
+       "type": "WarmupDecayLR",
+       "params": {
+         "total_num_steps": "auto",
+         "warmup_min_lr": "auto",
+         "warmup_max_lr": "auto",
+         "warmup_num_steps": "auto"
+        }
+    },
+    "zero_optimization": {
+        "stage": 2,
+        "allgather_partitions": true,
+        "allgather_bucket_size": 2e8,
+        "overlap_comm": true,
+        "reduce_scatter": true,
+        "reduce_bucket_size": 2e8,
+        "contiguous_gradients": true
+    },
+    "gradient_accumulation_steps": "auto",
+    "gradient_clipping": "auto",
+    "steps_per_print": 1e5,
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+}

configs/ds_configs/stage2_no_offloading_accelerate.conf

@@ -0,0 +1,20 @@
+{
+    "bf16": {
+        "enabled": "auto"
+    },
+    "zero_optimization": {
+        "stage": 2,
+        "allgather_partitions": true,
+        "allgather_bucket_size": 2e8,
+        "overlap_comm": true,
+        "reduce_scatter": true,
+        "reduce_bucket_size": "auto",
+        "contiguous_gradients": true
+    },
+    "gradient_accumulation_steps": "auto",
+    "gradient_clipping": "auto",
+    "steps_per_print": 1e5,
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+}

configs/ds_configs/stage3_no_offloading.conf

@@ -0,0 +1,41 @@
+{
+    "bf16": {
+        "enabled": "auto"
+    },
+    "optimizer": {
+        "type": "AdamW",
+        "params": {
+            "lr": "auto",
+            "betas": "auto",
+            "eps": "auto",
+            "weight_decay": "auto"
+        }
+    },
+    "scheduler": {
+       "type": "WarmupDecayLR",
+       "params": {
+         "total_num_steps": "auto",
+         "warmup_min_lr": "auto",
+         "warmup_max_lr": "auto",
+         "warmup_num_steps": "auto"
+        }
+    },
+    "zero_optimization": {
+        "stage": 3,
+        "overlap_comm": true,
+        "contiguous_gradients": true,
+        "sub_group_size": 1e9,
+        "reduce_bucket_size": "auto",
+        "stage3_prefetch_bucket_size": "auto",
+        "stage3_param_persistence_threshold": "auto",
+        "stage3_max_live_parameters": 1e9,
+        "stage3_max_reuse_distance": 1e9,
+        "stage3_gather_16bit_weights_on_model_save": true
+    },
+    "gradient_accumulation_steps": "auto",
+    "gradient_clipping": "auto",
+    "steps_per_print": 1e5,
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+}