GitHub - waltonfuture/Diff-eRank: [NeurIPS 2024] A Novel Rank-Based Metric for Evaluating Large Language Models (original) (raw)

Diff-eRank: A Novel Rank-Based Metric for Evaluating Large Language Models (NeurIPS 2024)

Lai Wei *, Zhiquan Tan *, Chenghai Li, Jindong Wang, Weiran Huang (*Equal Contribution).

Shanghai Jiao Tong University & Tsinghua University & William and Mary

Introduction

We introduce a rank-based metric called Diff-eRank, which is rooted in information theory and geometry principles. Diff-eRank evaluates LLMs by examining their hidden representations to quantify how LLMs discard redundant information after training. Specifically, we demonstrate its applicability in both single-modal (language) and multi-modal settings. For language models, our findings reveal that the Diff-eRank increases when the model scales up, which also demonstrates a consistent relationship with traditional metrics like loss and accuracy. For multi-modal models, we also propose an evaluation method based on rank for assessing alignment quality and we find that modern multi-modal large language models exhibit good alignment performance.

Calculation of Diff-eRank for LLMs

Setup

pip install transformers torch datasets

Calculation

from transformers import AutoTokenizer, AutoModel, AutoConfig import torch import math

R input N*d

def normalize(R): with torch.no_grad(): mean = R.mean(dim=0) R = R - mean norms = torch.norm(R, p=2, dim=1, keepdim=True) R = R/norms return R

def cal_cov(R): with torch.no_grad(): Z = torch.nn.functional.normalize(R, dim=1) A = torch.matmul(Z.T, Z)/Z.shape[0] return A

def cal_erank(A): with torch.no_grad(): eig_val = torch.svd(A / torch.trace(A))[1] entropy = - (eig_val * torch.log(eig_val)).nansum().item() erank = math.exp(entropy) return erank

def compute(R): return cal_erank(cal_cov(normalize(R)))

model_path = "facebook/opt-1.3b" # for example tokenizer = AutoTokenizer.from_pretrained(model_path) model = AutoModel.from_pretrained(model_path).cuda() config = AutoConfig.from_pretrained(model_path) untrained_model = AutoModel.from_config(config).to('cuda')

text = "We introduce a rank-based metric called Diff-eRank, which is rooted in information theory and geometry principles. Diff-eRank evaluates LLMs by examining their hidden representations to quantify how LLMs discard redundant information after training." # for example inputs = tokenizer(text, return_tensors="pt").to('cuda') with torch.no_grad(): R1 = model(inputs.input_ids)[0][0, :, :] R2 = untrained_model(inputs.input_ids)[0][0, :, :] erank1 = compute(R1) erank2 = compute(R2) RD = erank2 - erank1 print(RD)

Diff-eRank of Single Sentence

cd utils

python diff_erank_single_sentence.py

Diff-eRank of Dataset

Please download the datasets of wiki-en, dolly-15k, openwebtext2, hh-rlhf in huggingface and edit the data path in your scripts.

cd utils

python diff_erank_dataset.py

Calculate eRank for MLLMs

We provide an example script to calculate eRank for Qwen2.5-VL. Please check it in utils/erank-qwen2_5_vl.py

Citation

If you're using Diff-eRank in your research or applications, please cite using this BibTeX:

@inproceedings{weidiff, title={Diff-eRank: A Novel Rank-Based Metric for Evaluating Large Language Models}, author={Wei, Lai and Tan, Zhiquan and Li, Chenghai and Wang, Jindong and Huang, Weiran}, booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems}, year={2024} }