Language-Specific-Neurons

This repository is the official implementation of our paper Language-Specific Neurons: The Key to Multilingual Capabilities in Large Language Models in ACL 2024.

Language Neurons Found by LAPE

We provide our found language-specific neurons in LLaMA-2 (7B), LLaMA-2 (13B), LLaMA-2 (70B), BLOOM (7B), OPT (6.7B), Mistral (7B), and Phi-2 (2.7B).

You should use torch.load to load xxx.neuron.pth, each of which is a List[List[LongTensor]], neuron[i][j] represents the neuron indice of the i-th language in the j-th layer in the model. The language 0-6 indice stand for en, zh, fr, es, vi, id, ja. For example, LLaMA-2-7B[1][4]=tensor([6147, 9114, 9292]), which means that the Chinese neurons inside the 4-th layer of LLaMA-2-7B are of the indice 6147, 9114, and 9292.

Identifying Language-specific Neurons

Please use vllm==0.2.7 to run our code.

Record the activation state:

CUDA_VISIVLE_DEVICES=0 python activation.py -m meta-llama/Llama-2-7b-hf -l xx

Identifying language-specific neurons:

python identify.py

Computing PPL when Deactivating Neurons

You should first download the wikipedia texts from https://huggingface.co/datasets/wikimedia/wikipedia. Then tokenize them, concateneate them into a long list, and save them as a LongTensor in data/id.{lang}.train.llama. Finally, run the following code:

CUDA_VISIVLE_DEVICES=0 python ppl.py -m meta-llama/Llama-2-7b-hf -a activation_mask/xxx

Open-ended Generation when Deactivating Neurons

CUDA_VISIVLE_DEVICES=0 python generation.py -m meta-llama/Llama-2-7b-hf -a activation_mask/xxx