repository·main·Indexed 25 days ago
https://github.com/zjunlp/easyeditA unified framework for knowledge and model editing of Large Language Models (LLMs). EasyEdit enables the control and modification of LLM knowledge and behavior through insertion, updating, and erasing. It includes implementations such as SafeEdit for toxic region editing, DeCo (Dynamic Correction Decoding) and DoLa (Decoding by Contrasting Layers) for hallucination mitigation and steering in LLMs and MLLMs, and tools for identifying and manipulating knowledge neurons in BERT, GPT2, and GPT-Neo.
EasyEdit is a Python package designed to efficiently edit the behavior of Large Language Models (LLMs) like GPT-J, Llama, GPT-NEO, GPT2, and T5 (supporting models from 1B to 65B) within specific domains without degrading general performance.
The framework is organized into three core components:
BaseEditor for Factual Knowledge and Generation, or MultiModalEditor for Multimodal Knowledge).ROME, MEND).Reliability, Generalization, Locality, and Portability.EasyEdit2 implements several categories of steering methods to control LLM behavior:
Prompting primes a frozen pretrained model for a specific downstream task using a text prompt. There are two main categories:
Soft prompt methods included in PEFT are prompt tuning, prefix tuning, and P-tuning.
EasyEdit provides implementations for several knowledge editing scenarios:
To create steering vectors for a language model, follow these steps:
hparams/Steer/caa_hparams/generate_caa.yaml). For CAA, use:alg_name: caa
layers: [17]
multiple_choice: falsehparams/Steer/vector_generate.yaml to specify the model_name_or_path, dtype, device, and the paths for steer_train_hparam_paths, steer_train_dataset, and steer_vector_output_dir.question, matching, and not_matching keys, or use prepare_train_datasets(top_cfg).BaseVectorGenerator with your configuration and call generate_vectors(datasets).Vectors are saved to: {steer_vector_output_dir}/{steer_train_dataset}/{method_name}_vector.
# Example workflow for generating vectors
vector_generator = BaseVectorGenerator(top_cfg)
vector_generator.generate_vectors(datasets)To evaluate Concept Consistency (semantic similarity of generated concept definitions), follow these steps:
run_concept_editing.py, uncomment line 113: concept_consistency = True.easyeditor/editors/concept_editor.py, modify line 184 to set test_concept_consistency=concept_consistency.python examples/conceptedit_transform_check.py --method FT --model llama2chat --module intraWhen working with datasets where text is already split into words (like BioNLP2004), you must use is_split_into_words=True in the tokenizer. You also need to align the labels with the subword tokens generated by the tokenizer.
Use tokenizer.word_ids(batch_index=i) to map subword tokens back to their original word indices. Set labels for special tokens to -100 to ignore them during training.
def tokenize_and_align_labels(examples):
tokenized_inputs = tokenizer(examples["tokens"], truncation=True, is_split_into_words=True)
labels = []
for i, label in enumerate(examples[f"tags"]):
word_ids = tokenized_inputs.word_ids(batch_index=i)
previous_word_idx = None
label_ids = []
for word_idx in word_ids:
if word_idx is None:
label_ids.append(-100)
elif word_idx != previous_word_idx:
label_ids.append(label[word_idx])
else:
label_ids.append(-100)
previous_word_idx = word_idx
labels.append(label_ids)
tokenized_inputs["labels"] = labels
return tokenized_inputsTo use FSDP for distributed training of large models, you must first create a configuration file using the accelerate config command. This file defines your sharding strategy, CPU offloading settings, and auto-wrap policies.
Run the following command to start the interactive configuration process and save it to a specific file:
accelerate config --config_file fsdp_config.yamlKey configuration arguments to consider:
Sharding Strategy:[1] FULL_SHARD: Shards optimizer states, gradients, and parameters.[2] SHARD_GRAD_OP: Shards optimizer states and gradients.[3] NO_SHARDOffload Params: Determines whether to offload parameters and gradients to the CPU.Auto Wrap Policy:[1] TRANSFORMER_BASED_WRAP: Wraps layers based on specific class names.[2] SIZE_BASED_WRAP: Wraps layers based on parameter count.[3] NO_WRAPTransformer Layer Class to Wrap: A comma-separated string of case-sensitive transformer layer class names (e.g., T5Block, BertLayer) used when TRANSFORMER_BASED_WRAP is selected.Backward Prefetch: [1] BACKWARD_PRE, [2] BACKWARD_POST, or [3] NO_PREFETCH.State Dict Type: [1] FULL_STATE_DICT, [2] LOCAL_STATE_DICT, or [3] SHARDED_STATE_DICT.Before running, ensure the following directories exist and are properly configured: ./data, ./hparams, and ./hugging_cache.
Move run_concept_editing.py to the root directory (./) before execution.
Supported editing methods include: FT, ROME, MEMIT, and PROMPT.
Note for Llama2 users: If using LlaMA2-13B-Chat instead of LlaMA2-13B-Base, you must manually update the model_name in the corresponding .yaml file (e.g., ./hparams/[METHOD]/llama-7b.yaml) or provide your own configuration file.
python run_concept_editing.py --editing_method=ROME --edited_model gptj --hparams_dir=./hparams/ROME/gpt-j-6B --interTo use P-tuning for sequence classification, ensure you have peft, transformers, datasets, and evaluate installed.
!pip install -q peft transformers datasets evaluate