EasyEdit

https://github.com/zjunlp/easyedit
Admin
EasyEdit is a framework for knowledge editing in large language models that enables efficient...
Tokens:289,321
Snippets:2,968
Trust Score:10
License:MIT
Update:5 months ago
Show doc for...
Tokens:
Raw
### Install EasyEdit

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_ROME_llama.ipynb

Install the EasyEdit library to use its functionalities. This is a prerequisite for running the examples.

```bash
pip install "easyedit"
```

--------------------------------

### Install EasyEdit

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_ROME_llama.ipynb

Install the EasyEdit library to use its functionalities. This is a prerequisite for running the examples.

```bash
pip install "easyedit @ git+https://github.com/sapphir-ai/easyedit.git"
```

--------------------------------

### Install EasyEdit

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_ROME_llama.ipynb

Install the EasyEdit library to use its functionalities. This is a prerequisite for running the examples.

```bash
pip install "easyedit @ git+https://github.com/demonstrate/easyedit.git@main"
```

--------------------------------

### Install EasyEdit

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_ROME_llama.ipynb

Install the EasyEdit library using pip. This is a prerequisite for running the examples.

```bash
pip install easyedit
```

--------------------------------

### Full Training Example

Source: https://github.com/zjunlp/easyedit/blob/main/examples/MMEdit.md

A complete example demonstrating the setup and execution of a training process using SERACMultimodalTrainingHparams, CaptionDataset, and MultimodalTrainer.

```python
training_hparams = SERACMultimodalTrainingHparams.from_hparams('hparams/TRAINING/SERAC/minigpt4.yaml')
train_ds = CaptionDataset('data/caption_train_edit.json', config=training_hparams)
eval_ds = CaptionDataset('data/caption_eval_edit.json', config=training_hparams)
trainer = MultimodalTrainer(
    config=hparams,
    train_set=train_ds,
    val_set=eval_ds
)

trainer.run()
```

--------------------------------

### Install necessary libraries

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/task_guides/clm-prompt-tuning.mdx

Installs the required libraries for PEFT, transformers, and datasets. Use this command before starting the setup.

```bash
!pip install -q peft transformers datasets
```

--------------------------------

### Install EasyEdit

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_ROME_llama.ipynb

Install the EasyEdit library to enable knowledge editing functionalities. This is a prerequisite for running the examples.

```bash
pip install "easyedit @ git+https://github.com/ds4dp/easyedit.git@main"
```

--------------------------------

### Prepare Runtime Environment

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_ROME.ipynb

This section outlines the initial setup required for the EasyEdit library, including necessary installations and environment configurations.

```python
import os
import sys
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Set environment variables for distributed training if needed
# os.environ["WORLD_SIZE"] = "1"
# os.environ["RANK"] = "0"
# os.environ["LOCAL_RANK"] = "0"
# os.environ["NODE_RANK"] = "0"
# os.environ["MASTER_ADDR"] = "localhost"
# os.environ["MASTER_PORT"] = "29500"

# Add project root to Python path
# sys.path.append(os.path.abspath(os.path.join(os.getcwd(), "../../")))

# Import EasyEdit library
# from easyedit import EditingModel

# Define model and tokenizer names
MODEL_NAME = "gpt2-medium"

# Load model and tokenizer
# model = AutoModelForCausalLM.from_pretrained(MODEL_NAME)
# tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)

# Optional: Move model to GPU if available
# if torch.cuda.is_available():
#     model = model.cuda()

# print(f"Model {MODEL_NAME} loaded successfully.")

```

--------------------------------

### Install EasyEdit

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_ROME_InternLM.ipynb

Install the EasyEdit library to enable knowledge editing functionalities. This is a prerequisite for running the examples.

```python
import sys
import os

# Add the parent directory to the Python path
sys.path.append(os.path.abspath(os.path.join(os.getcwd(), '..')))

```

--------------------------------

### Setup Environment for EasyEdit

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_US_President.ipynb

Clone the EasyEdit repository, create a Python 3.9 environment using conda, and install the required libraries by running the provided commands in the terminal.

```bash
## Clone Repo
git clone https://github.com/zjunlp/EasyEdit.git
cd EasyEdit

## Create Environment
conda create -n EasyEdit python=3.9
conda activate EasyEdit
pip install -r requirements.txt
```

--------------------------------

### Environment Setup for Steering Experiments

Source: https://github.com/zjunlp/easyedit/blob/main/examples/SPLIT.md

Clone the repository, create a Conda environment, activate it, and install dependencies.

```bash
git clone https://github.com/zjunlp/EasyEdit.git
conda create -n spilt python=3.10
conda activate spilt
pip install -r requirements_2.txt
```

--------------------------------

### Install Watchdog for Previewing

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/README.md

Installs the watchdog module, which is required for previewing the documentation.

```bash
pip install watchdog
```

--------------------------------

### Install Dependencies

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/task_guides/dreambooth_lora.mdx

Install the required libraries for training, including PEFT from source. Ensure you have the latest version for compatibility.

```bash
pip install -r requirements.txt
pip install git+https://github.com/huggingface/peft
```

--------------------------------

### Fetch and Prepare Example Image

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/task_guides/image_classification_lora.mdx

Fetches an example image from a URL and prepares it for model inference. The image is converted to RGB format.

```python
from PIL import Image
import requests

url = "https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/beignets.jpeg"
image = Image.open(requests.get(url, stream=True).raw)
image
```

```python
image_processor = AutoImageProcessor.from_pretrained(repo_name)
encoding = image_processor(image.convert("RGB"), return_tensors="pt")
```

--------------------------------

### Setup Conda Environment for WISE

Source: https://github.com/zjunlp/easyedit/blob/main/examples/WISE.md

Create and activate a conda environment with Python 3.9 and install dependencies. This is the initial setup required to run WISE experiments.

```bash
conda create -n lifelong_edit python=3.9

conda activate lifelong_edit

pip install -r requirements.txt #https://github.com/zjunlp/EasyEdit/blob/main/requirements.txt
```

--------------------------------

### Install necessary libraries

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/task_guides/ptuning-seq-classification.mdx

Installs the required libraries for PEFT, Transformers, Datasets, and Evaluate.

```bash
!pip install -q peft transformers datasets evaluate
```

--------------------------------

### Install Python and Dependencies

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_QLoRA_llama3.2-3b.ipynb

Install a specific Python version (3.9) and update alternatives, then install pip and project requirements. Ensure the correct Python environment is set up before proceeding.

```python
!apt-get install python3.9
!sudo update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.9 1
!sudo update-alternatives --config python3
!apt-get install python3-pip
%pip install -r requirements.txt
```

--------------------------------

### Install Documentation Build Dependencies

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/README.md

Installs necessary packages for building documentation from the root of the repository.

```bash
pip install -e "<b>[docs]</b>"
```

--------------------------------

### Install Pip and Project Requirements

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_Multimodal_IKE.ipynb

Installs the Python package installer (pip) and then installs all project dependencies listed in the requirements.txt file.

```bash
!apt-get install python3-pip
!pip install -r requirements.txt
```

--------------------------------

### Environment Setup for EasyEdit

Source: https://github.com/zjunlp/easyedit/blob/main/examples/PersonalityEdit.md

Installs conda, creates a virtual environment with Python 3.9.7, activates it, and installs project requirements.

```shell
git clone https://github.com/zjunlp/EasyEdit.git
conda create -n EasyEdit python=3.9.7
...
conda activate EasyEdit
pip install -r requirements.txt
```

--------------------------------

### Example DeepSpeed Configuration File

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/accelerate/deepspeed-zero3-offload.mdx

An example YAML configuration file for DeepSpeed with ZeRO Stage-3 and CPU offloading for both optimizer and parameters. This setup is suitable for training large models.

```yaml
compute_environment: LOCAL_MACHINE
deepspeed_config:
  gradient_accumulation_steps: 1
  gradient_clipping: 1.0
  offload_optimizer_device: cpu
  offload_param_device: cpu
  zero3_init_flag: true
  zero3_save_16bit_model: true
  zero_stage: 3
distributed_type: DEEPSPEED
downcast_bf16: 'no'
dynamo_backend: 'NO'
fsdp_config: {}
machine_rank: 0
main_training_function: main
megatron_lm_config: {}
mixed_precision: 'no'
num_machines: 1
num_processes: 1
rdzv_backend: static
same_network: true
use_cpu: false
```

--------------------------------

### Install Documentation Builder Tool

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/README.md

Installs the special tool required for building the documentation.

```bash
pip install git+https://github.com/huggingface/doc-builder
```

--------------------------------

### Install Dependencies for LoRA Fine-tuning

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/task_guides/image_classification_lora.mdx

Install the necessary libraries for model training, including transformers, accelerate, evaluate, datasets, loralib, and peft. Use the `-q` flag for a quiet installation.

```bash
!pip install transformers accelerate evaluate datasets loralib peft -q
```

--------------------------------

### Install Dependencies

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/task_guides/int8-asr.mdx

Install necessary libraries for PEFT, Transformers, Datasets, Accelerate, Evaluate, and bitsandbytes.

```bash
!pip install -q peft transformers datasets accelerate evaluate jiwer bitsandbytes
```

--------------------------------

### Run PROMPT Example

Source: https://github.com/zjunlp/easyedit/blob/main/examples/PROMPT.md

Example command to run the PROMPT editing method. The `train_data_path` must be specified even though PROMPT does not use a training dataset.

```shell
python run_knowedit_llama2.py \
    --editing_method=IKE \
    --hparams_dir=./hparams/IKE/llama3-8b.yaml \
    --data_dir=./data/KnowEdit \
    --train_data_path=' '
```

--------------------------------

### Initialize and Start Trainer

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/task_guides/ptuning-seq-classification.mdx

Instantiate the Trainer with the PEFT model, training arguments, datasets, tokenizer, data collator, and evaluation function. Then, start the training process.

```python
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_datasets["train"],
    eval_dataset=tokenized_datasets["test"],
    tokenizer=tokenizer,
    data_collator=data_collator,
    compute_metrics=compute_metrics,
)

trainer.train()
```

--------------------------------

### Install EasyEdit2 Requirements

Source: https://context7.com/zjunlp/easyedit/llms.txt

Set up the Python environment and install necessary packages for EasyEdit2.

```bash
conda create -n easyedit2 python=3.10
conda activate easyedit2
pip install -r requirements_2.txt
```

--------------------------------

### Install Fairscale Library

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_IKE_InternLM.ipynb

Use this command to install the fairscale library, which is a dependency for EasyEdit.

```bash
!pip install fairscale
```

--------------------------------

### Example Evaluation Command

Source: https://github.com/zjunlp/easyedit/blob/main/README_2.md

An example of running the evaluation script with specific dataset path, evaluation methods, and model name.

```bash
python steer/evaluate/evaluate.py --generation_dataset_path results/my_dataset_results.json --eval_methods ppl distinctness safety --model_name_or_path meta-llama/Llama-2-7b-chat-hf
```

--------------------------------

### Full Evaluation Example

Source: https://github.com/zjunlp/easyedit/blob/main/examples/MMEdit.md

An example for setting up and running an evaluation process. Note that for evaluation, both `train_set` and `val_set` can be set to the evaluation dataset.

```python
hparams = SERACMultimodalHparams.from_hparams('hparams/SERAC/minigpt4.yaml')
# train_ds = CaptionDataset('data/caption_train_edit.json', config=hparams)
eval_ds = CaptionDataset('data/caption_eval_edit.json', config=hparams)
trainer = MultimodalTrainer(
    config=hparams,
    train_set=eval_ds,
    val_set=eval_ds
)

trainer.run()
```

--------------------------------

### Install PEFT from GitHub

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/install.mdx

Install the latest development version of PEFT directly from its GitHub repository. This is useful for trying new features or if you need a version not yet released on PyPI.

```bash
pip install git+https://github.com/huggingface/peft
```

--------------------------------

### Install Project Requirements

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_IKE_InternLM.ipynb

Installs all necessary Python packages listed in the requirements.txt file. Ensure you are in the EasyEdit directory before running.

```bash
!pip install -r requirements.txt
```

--------------------------------

### Run IKE Example

Source: https://github.com/zjunlp/easyedit/blob/main/examples/PROMPT.md

Example command to run the IKE editing method. The `train_data_path` is required for IKE to retrieve In-Context prompts.

```shell
python run_knowedit_llama2.py \
    --editing_method=IKE \
    --hparams_dir=./hparams/IKE/llama3-8b.yaml \
    --data_dir=./data/KnowEdit \
    --train_data_path='the_train_data_path'
```

--------------------------------

### Load and Use EasyEdit with Qwen

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_ROME_qwen.ipynb

Demonstrates loading a pre-trained Qwen model and tokenizer, then initializing EasyEdit for further operations.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Qwen/Qwen1.5-0.5B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype='auto', device_map="cuda")
```

--------------------------------

### Preview Documentation

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/README.md

Starts a local server to preview the documentation. The docs will be viewable at http://localhost:3000. Note that this command only works with existing doc files; new files require updating _toctree.yml and restarting the preview command.

```bash
doc-builder preview peft docs/source
```

--------------------------------

### Set up SteerEval Environment

Source: https://github.com/zjunlp/easyedit/blob/main/examples/SteerEval.md

Clone the repository, create and activate a conda environment, and install dependencies.

```bash
git clone https://github.com/zjunlp/EasyEdit.git
conda create -n steereval python=3.10
conda activate steereval
pip install -r requirements_2.txt
```

--------------------------------

### Prepare and Generate with EasyEdit (Second Example)

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_IKE_InternLM.ipynb

Sets up a second set of prompts for knowledge editing, tokenizes them, and generates model outputs. This demonstrates the flexibility of EasyEdit with different factual statements.

```python
ike_generation_prompts = [
    "Lionel Messi plays basketball. \
    Q: What is Messi's profession? A: Basketball. \
    Lionel Messi is a basketball player. \
    Q: What sports is Messi good at? A:",
    "The president of the US is Biden. \
    Q: Who is the president of the US? A: Biden. \
    Biden is the leader of the United State. \
    Q: The leader of the United State is? A:",
]
generation_prompts = [
    "Q: What sports is Messi good at? A:",
    "Q: The leader of the United State is? A:",
]

batch = tokenizer(generation_prompts, return_tensors='pt', padding=True)
edited_batch = tokenizer(ike_generation_prompts, return_tensors='pt', padding=True)



pre_edit_outputs = model.generate(
    input_ids=batch['input_ids'].to(model.device),
    attention_mask=batch['attention_mask'].to(model.device),
    max_new_tokens=15
)
post_edit_outputs = edited_model.generate(
    input_ids=edited_batch['input_ids'].to(edited_model.device),
    attention_mask=edited_batch['attention_mask'].to(edited_model.device),
    max_new_tokens=15
)
print('*'*100)

generation_max_length = batch['input_ids'].shape[-1]
edited_max_length = edited_batch['input_ids'].shape[-1]
for i in range(len(ike_generation_prompts)):
    print(f'Pre-Edit  Output: {tokenizer.decode( pre_edit_outputs[i][generation_max_length :], skip_special_tokens=True)}')
    print(f'Post-Edit Output: {tokenizer.decode(post_edit_outputs[i][edited_max_length :], skip_special_tokens=True)}')
    print('--'*50 )
```

--------------------------------

### Getting Started with PEFT and LoRA

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/README.md

This snippet demonstrates how to initialize a PEFT model using LoRA configuration for sequence-to-sequence language models. It loads a pre-trained model and tokenizer, applies a LoRA configuration, and then prints the trainable parameters.

```python
from transformers import AutoModelForSeq2SeqLM
from peft import get_peft_config, get_peft_model, LoraConfig, TaskType
model_name_or_path = "bigscience/mt0-large"
tokenizer_name_or_path = "bigscience/mt0-large"

peft_config = LoraConfig(
    task_type=TaskType.SEQ_2_SEQ_LM, inference_mode=False, r=8, lora_alpha=32, lora_dropout=0.1
)

model = AutoModelForSeq2SeqLM.from_pretrained(model_name_or_path)
model = get_peft_model(model, peft_config)
model.print_trainable_parameters()
# output: trainable params: 2359296 || all params: 1231940608 || trainable%: 0.19151053100118282
```

--------------------------------

### Load Configuration and Initialize

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit2_Example_CAA_sentiment.ipynb

Loads the configuration from a YAML file and sets up initial parameters for the model and device. Ensure the config file path is correct.

```python
import sys
sys.path.append('../')
from omegaconf import OmegaConf, DictConfig
from steer.vector_generators.vector_generators import BaseVectorGenerator
from steer.datasets import prepare_train_dataset
from steer.vector_appliers.vector_applier import BaseVectorApplier
from steer.datasets import prepare_generation_datasets
# model_path="../model/DeepSeek-R1-Distill-Qwen-7B"

top_cfg = OmegaConf.load("./config_sentiment.yaml")
# top_cfg.model_name_or_path = model_path
# top_cfg.device = "cuda:1"
top_cfg
```

--------------------------------

### List Installed Packages

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_Multimodal_IKE.ipynb

Displays a list of all Python packages currently installed in the environment. Useful for verifying installations.

```bash
!pip list
```

--------------------------------

### Setup model and training parameters

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/task_guides/seq2seq-prefix-tuning.mdx

Initializes the model, tokenizer, and defines hyperparameters for training. Sets environment variables for tokenizer parallelism and CUDA device.

```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, default_data_collator, get_linear_schedule_with_warmup
from peft import get_peft_config, get_peft_model, get_peft_model_state_dict, PrefixTuningConfig, TaskType
from datasets import load_dataset
from torch.utils.data import DataLoader
from tqdm import tqdm
import os

os.environ["TOKENIZERS_PARALLELISM"] = "false"
os.environ["CUDA_VISIBLE_DEVICES"] = "3"

device = "cuda"
model_name_or_path = "t5-large"
tokenizer_name_or_path = "t5-large"

text_column = "sentence"
label_column = "text_label"
max_length = 128
lr = 1e-2
num_epochs = 5
batch_size = 8
```

--------------------------------

### Install Dependencies

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_IKE_Baichuan.ipynb

Install Python 3.9 and necessary packages including python3-pip and requirements.txt. Also install the sentence-transformers library.

```bash
!apt-get install python3.9
!sudo update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.9 1
!sudo update-alternatives --config python3
!apt-get install python3-pip
!pip install -r requirements.txt
!pip install sentence_transformers
```

--------------------------------

### Pinyin Notation Example (JSON)

Source: https://github.com/zjunlp/easyedit/blob/main/examples/CKnowEdit.md

An example of a 'Pinyin Notation' data sample from CKnowEdit, demonstrating a prompt for pinyin, incorrect/correct responses, and portability/locality examples.

```json
{
    "prompt": "请给下面的字注音：莘庄",
    "target_old": "莘庄的拼音是：“shēn zhuāng”",
    "target_new": "莘庄的注音是：xīn zhuāng",
    "portability": [
        {
            "prompt": "请给下面的字注音：细莘",
            "answer": "细莘的注音是：xì xīn"
        }
    ],
    "locality": [
        {
            "prompt": "请给下面的字注音：莘莘学子",
            "answer": "莘莘学子的注音是：shēn shēn xué zǐ"
        }
    ],
    "rephrase": [
        "请问莘庄的正确拼音是什么？",
        "莘庄正确的注音怎么写？",
        "请注音：莘庄。",
        "莘庄的拼音应该怎么表示？"
    ]
}
```

--------------------------------

### Download Base Model (Qwen2.5-0.5B-Instruct)

Source: https://github.com/zjunlp/easyedit/blob/main/examples/CCKS2025.md

Command to download the Qwen2.5-0.5B-Instruct model using the huggingface-cli. The model will be saved to the specified local directory.

```shell
huggingface-cli download Qwen/Qwen2.5-0.5B-Instruct --local-dir hugging_cache/Qwen2.5-0.5B-Instruct
```

--------------------------------

### Install EasyEdit with Conda

Source: https://github.com/zjunlp/easyedit/blob/main/examples/ConceptEdit.md

Installs EasyEdit using conda, creating a dedicated Python 3.9.7 environment. Ensure you have conda installed and are using Python 3.9+.

```shell
git clone https://github.com/zjunlp/EasyEdit.git
conda create -n EasyEdit python=3.9.7
conda activate EasyEdit
pip install -r requirements.txt
```

--------------------------------

### Example FSDP Configuration File

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/accelerate/fsdp.mdx

This YAML file demonstrates a typical FSDP configuration, including sharding strategy, offload parameters, auto-wrap policy, and transformer layer class to wrap.

```yaml
command_file: null
commands: null
compute_environment: LOCAL_MACHINE
deepspeed_config: {}
distributed_type: FSDP
downcast_bf16: 'no'
dynamo_backend: 'NO'
fsdp_config:
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_backward_prefetch_policy: BACKWARD_PRE
  fsdp_offload_params: true
  fsdp_sharding_strategy: 1
  fsdp_state_dict_type: FULL_STATE_DICT
  fsdp_transformer_layer_cls_to_wrap: T5Block
gpu_ids: null
machine_rank: 0
main_process_ip: null
main_process_port: null
main_training_function: main
megatron_lm_config: {}
mixed_precision: 'no'
num_machines: 1
num_processes: 2
rdzv_backend: static
same_network: true
tpu_name: null
tpu_zone: null
use_cpu: false

```

--------------------------------

### Install SafeEdit Project Dependencies

Source: https://github.com/zjunlp/easyedit/blob/main/demo/SafeEdit_demo/README.md

Clone the repository, create and activate a Conda environment, and install project requirements. This includes installing dependencies for the main project and the SafeEdit demo.

```shell
git clone https://github.com/zjunlp/EasyEdit.git
conda create -n EasyEdit python=3.9.7
conda activate EasyEdit
cd EasyEdit
pip install -r requirements.txt
cd demo/SafeEdit_demo
pip install -r requirements.txt
```

--------------------------------

### Python Example Docstring with Accelerate

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/README.md

Demonstrates how to write an example docstring in Python using the `accelerate` library. Ensure the example is minimal, clear, and includes expected output for inference usage.

```python
>>> import time
>>> from accelerate import Accelerator
>>> accelerator = Accelerator()
>>> if accelerator.is_main_process:
...     time.sleep(2)
>>> else:
...     print("I'm waiting for the main process to finish its sleep...")
>>> accelerator.wait_for_everyone()
>>> # Should print on every process at the same time
>>> print("Everyone is here")
```

--------------------------------

### Setup Environment Variables and Model Parameters

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/task_guides/int8-asr.mdx

Configure the CUDA visible devices and define model, language, task, and dataset parameters for training.

```python
import os

os.environ["CUDA_VISIBLE_DEVICES"] = "0"
model_name_or_path = "openai/whisper-large-v2"
language = "Marathi"
language_abbr = "mr"
task = "transcribe"
dataset_name = "mozilla-foundation/common_voice_11_0"
```

--------------------------------

### Initialize Model and Tokenizer

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_IKE.ipynb

Sets up the tokenizer and language model for text generation. Ensure the model path is correct and the device is specified.

```python
tokenizer.pad_token_id = tokenizer.eos_token_id
tokenizer.padding_side='left'
device = 1
model = GPT2LMHeadModel.from_pretrained('./hugging_cache/gpt2-xl',trust_remote_code=True).to(f'cuda:{device}')
```

--------------------------------

### Prepare Runtime Environment

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_IKE.ipynb

This snippet is a placeholder for setting up the necessary runtime environment for the IKE method. Ensure all required libraries and dependencies are installed before proceeding.

```python
# Prepare the runtime environment
```

--------------------------------

### Example Output Comparison

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_KN_T5.ipynb

Illustrates the expected format for comparing pre-edit and post-edit model outputs for a given prompt.

```text
Output:
Prompt: Who was the designer of Lahti Town Hall?
Pre-Edit  Output: was the designer of Lahti Town Hall??? Who was the designer
Post-Edit Output: dendendenden
----------------------------------------------------------------------------------------------------
```

--------------------------------

### Install Additional Libraries

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_Multimodal_IKE.ipynb

Installs the matplotlib library for plotting and sentence_transformers for natural language processing tasks.

```bash
!pip install matplotlib
!pip install sentence_transformers
```

--------------------------------

### Set up Training Arguments

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/task_guides/ptuning-seq-classification.mdx

Configure training hyperparameters, evaluation strategy, and saving strategy using TrainingArguments.

```python
training_args = TrainingArguments(
    output_dir="your-name/roberta-large-peft-p-tuning",
    learning_rate=1e-3,
    per_device_train_batch_size=32,
    per_device_eval_batch_size=32,
    num_train_epochs=2,
    weight_decay=0.01,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
)
```

--------------------------------

### Install Python 3.9

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit_Example_Multimodal_IKE.ipynb

Installs Python 3.9, which may be required for specific dependencies or project compatibility.

```bash
!apt-get install python3.9
```

--------------------------------

### Load Configuration and Initialize

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit2_Example_CAA_translate.ipynb

Loads the configuration from a YAML file and sets up the necessary paths and parameters for the EasyEdit2 model. Ensure the config file exists and paths are correct.

```python
import sys
sys.path.append('..')
from omegaconf import OmegaConf, DictConfig
from steer.vector_generators.vector_generators import BaseVectorGenerator
from steer.datasets import prepare_train_dataset
from steer.vector_appliers.vector_applier import BaseVectorApplier
from steer.datasets import prepare_generation_datasets
# model_path="../model/Qwen2-7B-Instruct"

top_cfg = OmegaConf.load("./config_translate.yaml")
# top_cfg.model_name_or_path = model_path
# top_cfg.device = "cuda:0"
top_cfg
```

--------------------------------

### Install PEFT from PyPI

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/install.mdx

Use this command to install the latest stable version of PEFT from the Python Package Index.

```bash
pip install peft
```

--------------------------------

### Create FSDP Configuration File with Accelerate

Source: https://github.com/zjunlp/easyedit/blob/main/easyeditor/models/melo/peft_egg/docs/source/accelerate/fsdp.mdx

Run this command to create a FSDP configuration file using 🤗 Accelerate. Use the `--config_file` flag to specify a location; otherwise, it defaults to `default_config.yaml` in the Accelerate cache.

```bash
accelerate config --config_file fsdp_config.yaml
```

--------------------------------

### Install EasyEdit with Conda

Source: https://github.com/zjunlp/easyedit/blob/main/examples/MMEdit.md

Clone the repository and create a conda environment with Python 3.9+ for EasyEdit. Then, install the required packages.

```bash
git clone https://github.com/zjunlp/EasyEdit.git
conda create -n EasyEdit python=3.9.7
...
pip install -r requirements.txt
```

--------------------------------

### Load Configuration and Prepare Dataset

Source: https://github.com/zjunlp/easyedit/blob/main/tutorial-notebooks/EasyEdit2_Example_CAA_r1_control.ipynb

Loads a configuration file using OmegaConf and prepares the training dataset. Ensure the config file path is correct.

```python
import sys
sys.path.append('../')
from omegaconf import OmegaConf, DictConfig
from steer.vector_generators.vector_generators import BaseVectorGenerator
from steer.datasets import prepare_train_dataset
from steer.vector_appliers.vector_applier import BaseVectorApplier
from steer.datasets import prepare_generation_datasets
# model_path="../model/DeepSeek-R1-Distill-Qwen-7B"

top_cfg = OmegaConf.load("./config_r1_control.yaml")
# top_cfg.model_name_or_path = model_path
```

--------------------------------

### ZsRE Knowledge Editing Example

Source: https://github.com/zjunlp/easyedit/blob/main/examples/KnowEdit.md

This JSON structure represents a ZsRE (Zero-Shot Relation Editing) example. It includes subject, target, prompt, ground truth, and rephrased prompts for relation editing. It also contains locality and portability examples for more specific knowledge manipulation.

```json
{
        "subject": "Epaspidoceras",
        "target_new": "Noctuidae",
        "prompt": "Which family does Epaspidoceras belong to?",
        "ground_truth": [
            "Aspidoceratidae"
        ],
        "rephrase_prompt": "What family are Epaspidoceras?",
        "cond": "Geometridae >> Noctuidae || Which family does Epaspidoceras belong to?",
        "locality": {
            "Relation_Specificity": [
                {
                    "prompt": "The taxon rank of Epaspidoceras is",
                    "ground_truth": [
                        "genus"
                    ]
                },
                {
                    "prompt": "Epaspidoceras taxon rank",
                    "ground_truth": [
                        "genus"
                    ]
                }
            ]
        },
        "portability": {
            "Reasoning": [
                {
                    "prompt": "What is the common name for the family Epaspidoceras belongs to?",
                    "ground_truth": "Owlet moths"
                }
            ]
        }
}

```
Tokens:
Raw