Repository files navigation

Code for paperAligning with Human Judgement: The Role of Pairwise Preference in Large Language Model Evaluators

Link to paper:Aligning with Human Judgement: The Role of Pairwise Preference in Large Language Model Evaluators (arXiv preprint arXiv:2403.16950)
This paper has been accepted byCOLM 2024.

If you are interested in pairwise evaluator, please also checkout our latest work onzero-shot automatic prompt optimization for pairwise evaluators.

We provide a ready-to-use Python library for Pairwise preference ranking (PairS). We show a ranking demonstration below.For an input source text and a sequence of output candidates,PairsGreedy andPairsBeam can be used to rank the output candidates inascending order.We currently support the following base models:google/gemma-2-9b-it, google/gemma-2-27b-it, meta-llama/Meta-Llama-3-8B-Instruct, microsoft/Phi-3-medium-4k-instruct, microsoft/Phi-3-mini-4k-instruct, mistralai/Mistral-7B-Instruct-v0.1, meta-llama/Llama-2-7b-chat-hf, meta-llama/Llama-2-13b-chat-hf, HuggingFaceH4/zephyr-7b-beta, gpt-3.5-turbo, gpt-4-turbo.

frompairsimportPairsGreedy,PairsBeamfromscripts.utilsimportshuffle_lists,load_summEval# Load example datasumm_eval_path='data/SummEval/model_annotations.aligned.paired.jsonl'input_doc,output_doc,_=load_summEval(summ_eval_path,flat_output=False)doc_id=42input,output=input_doc[doc_id],output_doc[doc_id]input,output=shuffle_lists(input,output)# The same input source text corresponds to multiple output summariesprint('Number of summary candidates:',len(output))method='PairsGreedy'ifmethod=='PairsGreedy':# Set hyperparametersparams= {# 'engine': "mistralai/Mistral-7B-Instruct-v0.1",'engine':"meta-llama/Llama-2-7b-chat-hf",'api_call':0,'with_input':True,# Use the prompt template for task with context input, e.g. Summarization'calibrate':False,# For each pairwise comparison, we average the probabilities of both permutations to cancel the positional bias.    }# Rank the output summaries from low to high qualityindices=PairsGreedy(input[0],output,params)print(indices)elifmethod=='PairsBeam':# Set hyperparametersparams= {'engine':"mistralai/Mistral-7B-Instruct-v0.1",'beam_size':2000,'api_call':0,'prob_gap':0.1,'with_input':True,'calibrate':False,    }# Rank the output summaries from low to high qualityindices=PairsBeam(input[0],output,params)print(indices)

We also present the original code (in the folderscripts/) to evalute on the datasets reported in the paper.

For NewsRoom and SummEval

bashpairs_run.sh

We provide a Notebook demonstrations innotebooks/.

Load dataset: We put all datasets loading inscripts/utils.py.

Prompts: We put all prompts and instructions inscripts/prompts.py.

Base models: We supports the following base models,mistralai/Mistral-7B-Instruct-v0.1,meta-llama/Llama-2-7b-chat-hf, all versions ofGPT-3.5-turbo andGPT-4-turbo.

Hyper-parameters:

• dataset: We support 3 datasets, 'newsroom', 'SummEval' and 'hanna'.
• eval_method: For all PairS method, we use 'pairwise comparison'.
• engine: The base models.
• with_input: If the data format has input text. For example, the summarization task has source text as input, but story writing task has no input text.
• confidence_beam:True for PairS-beam andFalse for PairS-greedy.
• prob_gap: The uncertainty tolerance.$0.1$ represents we will create beam candidates for both A and B if$0.5-0.1 < P(A\succ B) < 0.5+0.1$.
• calibrate: LLMs suffer from positional bias. Set this asTrue will average the probabilities of both permutations of A and B for each pairwise comparison. This will cancel the positional bias.

More details and comments will be added soon.

The PairS-Greedy can be understood as a merge sort with pairwise comparison by LLMs, while the PairS-Beam is to do a beam-search for each merge operation. In order to improve the beam search efficiency and limit the search space, we also apply a local uncertainty-based prunning mechanism.

We show the algorithm of the modified merge operation for PairS-Beam below.

If you find our work helpful, please consider citing our paper:

@article{liu2024aligning,  title={Aligning with Human Judgement: The Role of Pairwise Preference in Large Language Model Evaluators},  author={Liu, Yinhong and Zhou, Han and Guo, Zhijiang and Shareghi, Ehsan and Vulic, Ivan and Korhonen, Anna and Collier, Nigel},  journal={arXiv preprint arXiv:2403.16950},  year={2024}}