Merge pull request #45 from PAIR-code/patchscopes-code

Patchscopes code

patchscopes/code/README.md

@@ -0,0 +1,50 @@
+##  🩺 Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models
+
+
+### Overview
+We propose a framework that decodes specific information from a representation within an LLM by “patching” it into the inference pass on a different prompt that has been designed to encourage the extraction of that information. A "Patchscope" is a configuration of our framework that can be viewed as an inspection tool geared towards a particular objective.
+
+For example, this figure shows a simple Patchscope for decoding what is encoded in the representation of "CEO" in the source prompt (left). We patch a target prompt (right) comprised of few-shot demonstrations of token repetitions, which encourages decoding the token identity given a hidden representation.
+
+[**[Paper]**](https://arxiv.org/abs/2401.06102) [**[Project Website]**](https://pair-code.github.io/interpretability/patchscopes/)
+
+<p align="left"><img width="60%" src="images/patchscopes.png" /></p>
+
+### 💾 Download textual data
+The script is provided [**here**](download_the_pile_text_data.py). Use the following command to run it:
+```python
+python3 download_the_pile_text_data.py
+```
+
+### 🦙 For using Vicuna-13B
+Run the following command for using the Vicuna 13b model (see also details [here](https://huggingface.co/CarperAI/stable-vicuna-13b-delta)):
+```python
+python3 apply_delta.py --base meta-llama/Llama-2-13b-hf --target ./stable-vicuna-13b --delta CarperAI/stable-vicuna-13b-delta
+```
+
+### 🧪 Experiments
+
+#### (1) Next Token Prediction
+The main code used appears [here](next_token_prediction.ipynb).
+#### (2) Attribute Extraction
+For this experiment, you should download the `preprocessed_data` directory.
+The main code used appears [here](attribute_extraction.ipynb).
+#### (3) Entity Processing
+The main code used appears [here](entity_processing.ipynb). The dataset is available for downloading [here](https://github.com/AlexTMallen/adaptive-retrieval/blob/main/data/popQA.tsv).
+#### (4) Cross-model Patching
+The main code used appears [here](patch_cross_model.ipynb).
+#### (5) Self-Correction in Multi-Hop Reasoning
+For this experiment, you should download the `preprocessed_data` directory.
+The main code used appears [here](multihop-CoT.ipynb). The code provided supports the Vicuna-13B model.
+
+### 📙 BibTeX
+```bibtex
+@misc{ghandeharioun2024patchscopes,
+      title={Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models},
+      author={Ghandeharioun, Asma and Caciularu, Avi and Pearce, Adam and Dixon, Lucas and Geva, Mor},
+      year={2024},
+      eprint={2401.06102},
+      archivePrefix={arXiv},
+      primaryClass={cs.CL}
+}
+```

patchscopes/code/apply_delta.py

@@ -0,0 +1,51 @@
+"""
+Usage:
+python3 apply_delta.py --base /path/to/model_weights/llama-13b --target stable-vicuna-13b --delta pvduy/stable-vicuna-13b-delta
+
+The code was adopted from https://github.com/GanjinZero/RRHF/blob/main/apply_delta.py
+"""
+import argparse
+
+import torch
+from tqdm import tqdm
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+
+def apply_delta(base_model_path, target_model_path, delta_path):
+    print("Loading base model")
+    base = AutoModelForCausalLM.from_pretrained(
+        base_model_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+
+    print("Loading delta")
+    delta = AutoModelForCausalLM.from_pretrained(delta_path, torch_dtype=torch.float16, low_cpu_mem_usage=True)
+    delta_tokenizer = AutoTokenizer.from_pretrained(delta_path)
+
+    DEFAULT_PAD_TOKEN = "[PAD]"
+    base_tokenizer = AutoTokenizer.from_pretrained(base_model_path, use_fast=False)
+    num_new_tokens = base_tokenizer.add_special_tokens(dict(pad_token=DEFAULT_PAD_TOKEN))
+
+    base.resize_token_embeddings(len(base_tokenizer))
+    input_embeddings = base.get_input_embeddings().weight.data
+    output_embeddings = base.get_output_embeddings().weight.data
+    input_embeddings[-num_new_tokens:] = 0
+    output_embeddings[-num_new_tokens:] = 0
+
+    print("Applying delta")
+    for name, param in tqdm(base.state_dict().items(), desc="Applying delta"):
+        assert name in delta.state_dict()
+        param.data += delta.state_dict()[name]
+
+    print("Saving target model")
+    base.save_pretrained(target_model_path)
+    delta_tokenizer.save_pretrained(target_model_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base-model-path", type=str, required=True)
+    parser.add_argument("--target-model-path", type=str, required=True)
+    parser.add_argument("--delta-path", type=str, required=True)
+
+    args = parser.parse_args()
+
+    apply_delta(args.base_model_path, args.target_model_path, args.delta_path)