Home

Awesome

RefChecker for Fine-grained Hallucination Detection

| 🔥 News | 🤖️ Demo | 🚀 Quick Start | 💾 Benchmark | 📖 Docs |

RefChecker provides a standardized assessment framework to identify subtle hallucinations present in the outputs of large language models (LLMs).

<p align="center"> <img src="imgs/framework.png" alt="RefChecker Framework" style="width:800px"> <br> <b>Figure</b>: RefChecker Framework </p>

🌟 Highlighted Features

You can explore RefChecker in the following ways:

<a id='news'></a>

🔥 News

❤️ Citation

Please check out the paper here: https://arxiv.org/pdf/2405.14486

If you use RefChecker in your work, please cite us:

@article{hu2024refchecker,
      title={RefChecker: Reference-based Fine-grained Hallucination Checker and Benchmark for Large Language Models}, 
      author={Xiangkun Hu and Dongyu Ru and Lin Qiu and Qipeng Guo and Tianhang Zhang and Yang Xu and Yun Luo and Pengfei Liu and Yue Zhang and Zheng Zhang},
      year={2024},
      eprint={2405.14486},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

<a id='demo_website'></a>

🤖️ Demo Website

<img src="imgs/demo.gif" alter="Demo" style="width:800px">

You can first setup a demo website and then use the web UI to try RefChecker as the animation shows above. There are four steps to perform hallucination detection in it:

  1. Extract Triplets: You can start with typing what you want to check in the top-left box. Then click the Next Step button on the right side. The checker will extract triplets in your text and show them in the bottom-left area.
  2. Gather Reference: You can then add reference text in the top-right box and click the Next Step button. If you don’t have reference text, leave the box empty and click the button anyway. We will retrieve some references with the text to be checked using search engines.
  3. Fact Checking: With the text to be checked and the retrieved reference text, the checker will perform fact checking then. The checking results will be shown in the bottom-left area, with ✅/❌/❓ indicating factual/hallucinatory/neutral. An overall factuality score will be given aside.
  4. Localization: You can then click the Next Step button and the checker will perform triplet localization. You can click the button on the left of each triplet to see the localization result.

<a id='quick_start'></a>

🚀 Quick Start

Setup Environment

First create a python environment using conda or virtualenv. Clone this repo and change path into the root directory. Then install:

pip install -e .
python -m spacy download en_core_web_sm

Install optional dependencies to use open source extractors (Mistral, Mixtral) or enable acceleration for RepCChecker.

pip install -e .[open-extractor,repcex]

Code Examples

Choose Models for the Extractor and Checker

We use litellm as to invoke the LLMs. Please check the document for how to setup the model for different LLM providers: https://docs.litellm.ai/docs/providers . We give some examples below:

Setup the enviroment variables if you are not using AWS EC2 instance

If you are using AWS EC2, make sure your region has the access to the model

export AWS_ACCESS_KEY_ID=<your_aws_access_key_id>
export AWS_SECRET_ACCESS_KEY=<your_aws_secret_access_key>
export AWS_REGION_NAME=<your_aws_region_name>

import os
from refchecker import LLMExtractor, LLMChecker

# Claude 3 Sonnet from Amazon Bedrock
model = 'bedrock/anthropic.claude-3-sonnet-20240229-v1:0'

extractor = LLMExtractor(model=model, batch_size=8)
checker = LLMChecker(model=model, batch_size=8)

You can also setup the enviroment variables in terminal to avoid disclosing these information in the code:

export AWS_ACCESS_KEY_ID=<your_aws_access_key_id>
export AWS_SECRET_ACCESS_KEY=<your_aws_secret_access_key>
export AWS_REGION_NAME=<your_aws_region_name>

import os
from refchecker import LLMExtractor, LLMChecker


os.environ["OPENAI_API_KEY"] = "<your_openai_api_key>"

# GPT-4o from OpenAI
model = 'gpt-4o'

extractor = LLMExtractor(model=model, batch_size=8)
checker = LLMChecker(model=model, batch_size=8)

Please use vllm to setup the API server for open source LLMs. For example, use the following command to deploy a Llama 3 8B hosted on HuggingFace:

python -m vllm.entrypoints.openai.api_server \
  --model meta-llama/Meta-Llama-3-8B-Instruct \
  --tensor-parallel-size 8 \
  --dtype auto \
  --api-key sk-123456789 \
  --gpu-memory-utilization 0.9 \
  --port 5000

Setup the api key:

export OPENAI_API_KEY=sk-123456789

Then we can initilize the extractor and checker with api_base:

import os
from refchecker import LLMExtractor, LLMChecker


# Note the prefix "openai/" here
model = "openai/meta-llama/Meta-Llama-3-8B-Instruct"
api_base = "http://0.0.0.0:5000/v1"
extractor = LLMExtractor(model=model, batch_size=8, api_base=api_base)
checker = LLMChecker(model=model, batch_size=8, api_base=api_base)

We fine-tuned a Mistral 7B model for claim extraction. Deploy it with vllm:

python -m vllm.entrypoints.openai.api_server \
  --model dongyru/Mistral-7B-Claim-Extractor \
  --tensor-parallel-size 8 \
  --dtype auto \
  --api-key sk-123456789 \
  --gpu-memory-utilization 0.9 \
  --port 5000

Then we can initilize the extractor as follows:

extractor = LLMExtractor(
  model="openai/dongyru/Mistral-7B-Claim-Extractor", 
  batch_size=8, 
  api_base="http://0.0.0.0:5000/v1"
)

We also offer non-LLM checker for efficent checking:

from refchecker import AlignScoreChecker, NLIChecker


# Details see paper: https://arxiv.org/abs/2305.16739
checker = AlignScoreChecker(device=0, batch_size=128)

# See https://huggingface.co/ynie/roberta-large-snli_mnli_fever_anli_R1_R2_R3-nli
checker = NLIChecker(device=0, batch_size=128)

Run Extraction and Checking

Both the extractor and checker takes a batch of inputs:

# Batch of questions (optional)
questions = ['question 1', 'question 2']
# Batch of model responses
responses = ['response 1', 'response 2']

extraction_results = extractor.extract(
    batch_responses=responses,
    batch_questions=questions,
    max_new_tokens=1000
)

batch_claims = [[c.content for c in res.claims] for  res in extraction_results]

references = ['reference 1', 'reference 2']
batch_labels = checker.check(
    batch_claims=batch_claims,
    batch_references=references,
    max_reference_segment_length=0
)

The extraction_results is a list of RCClaim objects defined in refchecker/base.py.

Command Line Interface

We provide a command-line interface to run RefChecker in a console:

usage: refchecker-cli [-h] --input_path INPUT_PATH --output_path OUTPUT_PATH
                     [--cache_dir CACHE_DIR]
                     [--extractor_name EXTRACTOR_NAME]
                     [--extractor_max_new_tokens EXTRACTOR_MAX_NEW_TOKENS]
                     [--claim_format {triplet, subsentence}]
                     [--checker_name CHECKER_NAME]
                     [--extractor_api_base EXTRACTOR_API_BASE]
                     [--checker_api_base CHECKER_API_BASE]
                     [--repc_classifier_name {svm,svm_ensemble,nn,nn_ensemble}]
                     [--retriever_name {google}]
                     [--aggregator_name {strict,soft,major}]
                     [--use_retrieval]
                     [--batch_size_extractor BATCH_SIZE_EXTRACTOR]
                     [--batch_size_checker BATCH_SIZE_CHECKER]
                     [{extract,check,extract-check}]

positional arguments:
  {extract,check,extract-check}
     extract:       Extract claims from provided responses.
     check:         Check whether the provided claims are factual.
     extract-check: Extract claims and check whether they are factual.

options:
  -h, --help  show this help message and exit
  --input_path INPUT_PATH
              Input path to the json file.
  --output_path OUTPUT_PATH
              Output path to the result json file.
  --cache_dir CACHE_DIR
              Path to the cache directory. Default: ./.cache.
  --extractor_name EXTRACTOR_NAME
              Model used for extracting claims. Default: bedrock/anthropic.claude-3-sonnet-20240229-v1:0
  --extractor_max_new_tokens EXTRACTOR_MAX_NEW_TOKENS
              Max generated tokens of the extractor, set a larger value for longer documents. Default: 500
  --claim_format {triplet, subsentence}
              The format of the extracted claims. Default: triplet
  --checker_name CHECKER_NAME
              Model used for checking whether the claims are factual. Default: bedrock/anthropic.claude-3-sonnet-20240229-v1:0
  --extractor_api_base EXTRACTOR_API_BASE
              API base URL if using vllm for deploying the extractor.
  --checker_api_base CHECKER_API_BASE
              API base URL if using vllm for deploying the checker
  --repc_classifier_name {svm,svm_ensemble,nn,nn_ensemble}
              Classifier Model used for RepC checker, only valid when RepC checker is used.
              Default: nn_ensemble, neural network classifier with layer ensemble.
  --retriever_name {google}
              Model used for retrieving reference (currently only google is supported).
              Default: google.
  --aggregator_name {strict,soft,major}
              Aggregator used for aggregating the results from multiple triplets.
              Default: soft.
              *  strict: If any of the triplets is Contradiction, the response is
              Contradiction. If all of the triplets are Entailment, the response is
              Entailment. Otherwise, the response is Neutral.
              *  soft:   The ratio of each category is calculated.
              *  major:  The category with the most votes is selected.
  --use_retriever  
              Whether to use retrieval to find the reference for checking. Required
              if the reference field in input data is not provided.
  --serper_api_key SERPER_API_KEY
              Path to the serper api key file. Required if the google retriever is
              used.
  --batch_size_extractor BATCH_SIZE_EXTRACTOR
              Batch size for batching inference of eatractor. Default: 16.
  --batch_size_checker BATCH_SIZE_CHECKER
              Batch size for batching inference of checker. Default: 16.

To extract claim triplets from LLM-generated responses, do:

refchecker-cli extract \
  --input_path {INPUT_PATH} \
  --output_path {OUTPUT_PATH} \
  --extractor_name {EXTRACTOR_NAME} \
  --extractor_api_base {EXTRACTOR_API_BASE}

The input json file contains a list of

{
   "response": "",  # required, the response to be checked
   "question": "",  # optional if the question is not important (e.g., in summarization)
   "reference": "", # required, the reference for checking
   ...
}

In the output json file, each item is added with a claims field, containing a list of [head, relation, tail].

To check hallucinations at triplet level, do:

refchecker-cli check \
  --input_path {INPUT_PATH} \
  --output_path {OUTPUT_PATH} \
  --checker_name {CHECKER_NAME} \
  --checker_api_base {CHECKER_API_BASE} \
  --aggregator_name {strict,soft,major}

The input json file contains a list of

{
   "response": "",  # required, the response to be checked
   "claims": [
       ["head1", "relation1", "tail1"],
       ["head2", "relation2", "tail2"],
       ...
   ]                # required, the corresponding triplets of the response
   "reference": "", # optional if a retriever is used to get reference
   ...
}

In the output json file, each item is added with the following fields:

{
   "Y": Union[str, dict],  # aggregated predictions on the whole response
   "ys": [
       "Entailment",
       "Neutral",
       "Contradiction",
       ...
   ]                # checker predictions on each triplet
   "reference": "", # added if a retriever is used to get reference
   ...
}

The format of aggregated predictions Y depends on the selected aggregator. It is a str as “Entailment”, “Neutral”, or “Contradiction” if strict or major aggregators are used. It is a dict containing ratios of each category if the soft aggregator is used. We additionally include a special category “Abstain” introduced in Evaluation Metric.

Note that the retriever is required in the zero-context setting, where no reference is provided by users. You can activate it by adding the --use_retriever flag and specifying --retriever_name. Currently we only support a google-based retriever. Feel free to try your own retrieval system and welcome to contribute.

For using the google retriever and/or the OpenAI models, you should provide corresponding API keys by specifying --serper_api_key and/or --openai_key.

Finally, you can use the whole extraction and checking pipeline by:

refchecker-cli extract-check \
  --input_path {INPUT_PATH} \
  --output_path {OUTPUT_PATH} \
  --extractor_name {EXTRACTOR_NAME} \
  --checker_name {CHECKER_NAME} \
  --extractor_api_base {EXTRACTOR_API_BASE} \
  --checker_api_base {CHECKER_API_BASE} \
  --aggregator_name {strict,soft,major} \
  <other optional flags>

You can try the command-line with example scripts and example input data here.

<a id='documentation'></a>

📖 Our Approach for Fine-grained Hallucination Detection

LLMs exhibit a susceptibility to generate hallucinated contents that can be challenging to discern, possibly leading users astray. To address this, this project centers on the construction of a standardized assessment framework aimed at identifying such hallucinations present in the outputs of LLMs. Specifically, we try to answer the the following questions:

Table of Contents

A Taxonomy of Hallucination

Hallucinations are claims made by LLMs not supported by factual knowledge, which we refer to as references; detecting hallucinations involves comparing the claims against the references. This process depends on tasks, contexts, granularity of checking and how we categorize them. We will discuss them in turn.

<a id='setting_definitions'></a>

Context, Tasks and References

When LLM generates a response, there is a great deal of variations as how contexts are provided which, in turn, determines where and how to identify the reference:

For both Noisy Context and Accurate Context tasks, we take the documents in the prompt as the references. The following figure illustrate the differences of the three settings.

<img src="imgs/settings.png" alt="Illustration of three settings of context, tasks and references." style="width:800px"> <p align="center">Illustration of three settings of context, tasks and references.</p>

Definition of Hallucination and Granularity of Checking

Hallucination detection is challenging because non-factualness often creeps in and corrupts parts of a response subtly. Rather than declaring the entire response as hallucinatory or not, in this work we inspect claims that are embedded in a response. With that, the basic relationship between a response and its corresponding references can be visualized as a Venn diagram shown below.

These three categories align closely with the concepts of support, refute, and not enough information within the fact-checking literature, and they are commonly used in Natural Language Inference (NLI).

<p align="center"><img src="imgs/venn.png" alt="Definition of Hallucinations." style="width:500px"></p> <p align="center">Definition of Hallucinations.</p>

Moving on, the remaining claims, represented by blue stars, contain information that is not addressed in the response. Whether this information should be incorporated into the response depends on the specific task at hand. In tasks like machine translation, omitting this information can result in a loss of information and should be considered a form of hallucination. In this release, we concentrate on question answering (QA) tasks, as well as text summarization and information extraction. In these tasks, missing facts are not regarded as hallucinations, their relevance may or may not be important. For example, in QA tasks the missing claims reflect retrieval quality instead; we leave this for future work.

<a id='granularity'></a>

The Granularity of Claim

Informally, claims are the unit for the checking. Previous works use sentences in the response as claims (SelfCheckGPT), or generate short phrases (i.e. sub-sentences) as the claims produced by LLM’s in-context learning (FActScore, FACTOOL). This work explores the approach of representing claims with knowledge triplets. This concept is inspired by the field of knowledge graph studies, where triplets are employed to encapsulate factual knowledge units. Knowledge triplets adopt a (subject, predicate, object) structure to capture fine-grained information within the response. We call the triplet represented claims as Claim-Triplets. Here is an example of model response and the claim-triplets extracted by Claude 2 (see Triplet Extraction for the details of claim-triplet extraction):

Richard Mulligan played Mr. Kincaid on The Partridge Family.

SubjectPredicateObject
Richard MulliganplayedMr. Kincaid
Mr. Kincaidchracter inThe Partridge Family

<a id='benchmark_dataset'></a>

Benchmark Dataset

We have assembled a benchmark dataset comprising 300 examples, with 100 examples for each of the three settings mentioned earlier. The examples are randomly sampled from the data sources listed in the following table:

SettingData SourceTaskReferences
Zero ContextNaturalQuestions (dev set)Closed Book QAAnnotated Long Answer
Noisy ContextMS MARCO (dev set)RAGRetrieved Passages
Accurate Contextdatabricks-dolly-15kSummarization, Closed QA, Information ExtractionInput Context

We collect responses from 7 LLMs on the benchmark: GPT4, GPT-3.5-Turbo, InstructGPT, Falcon (Falcon-40B-Instruct), Alpaca (Alpaca-7B), LLaMA2(70B-Chat) and Claude 2. The code and examples for response collection can be found in response_collection.

We performed a human evaluation of responses generated by seven LLMs on this benchmark dataset. The process is shown in the figure below which involved three steps: gathering responses, extracting claim-triplets, and asking human annotators to evaluate these claims. 23% of the claim-triplets were double annotated, with 95.0% Inter-Annotator Agreement. We employ Claude 2 for the claim-triplet extraction for human evaluations, but it can be replaced with any triplet extraction model. During the annotation, we also ask the annotators to mark low quality triplets. These identified low quality triplets are then removed from the benchmark dataset. We will release the data and results upon approval.

<img src="imgs/evaluation.png" alt="The evaluation process under the Zero Context setting as an example." style="width:600px"> <p align="center">The evaluation process under the Zero Context setting as an example.</p>

Evaluation Metric

We evaluate using the frequency of each hallucination label. For example, if there are 3 Entailment, 5 Neutral, and 2 Contradiction claims out of 10 in a specific response, the rates for Entailment, Neutral, and Contradiction in this response are 0.3, 0.5, and 0.2 respectively. The overall rates for each label are averaged across all responses in the benchmark to obtain the final evaluation metrics. Alternatively, we can aggregate at response level according to some rules, the most aggressive is to classify a response as hallucination if it contains at least one hallucinated claim triplet.

Sometimes, the LLM declines to answer questions by generating responses such as I'm sorry, I can't provide an accurate answer without more context. Could you please provide more information? When this happens, the claim-triplet extractor fails to get triplets from these responses. So, we introduce a new label called Abstain to for such cases. We treat these responses as individual claims, and in such cases, the Abstain Rate is 1 for that specific response. In practice, if the claim-triplet extractor doesn't give any triplets for a response, we treat this response as Abstain.

Automatic Hallucination Checker

Using automatic hallucination checker is a more scalable way for evaluating LLMs. Our automatic hallucination checker is a 3-stage pipeline, consisting of a knowledge extraction model $E$, a checker $C$, and aggregation rules $\tau$. Initially, we utilize $E$ to decompose the input text $T$ into a set of knowledge triplets $k_{1:N}$​. Each of these triplets undergoes verification by $C$. Subsequently, based on predefined rules $\tau$, the individual results $y_{1:N}$​ are aggregated to determine the overall hallucination label $Y$ for the given text. We delve into the specifics of each component in the subsequent parts and also elaborate on our strategies for enhancing the checker's performance through self-supervision.

<a id='extraction'></a>

Claim Extraction

Our checking framework hinges on a key assumption: the decomposition of the original text into triplets facilitates finer-grained detection and more accurate evaluation. The extraction of these triplets plays a pivotal role in achieving this objective. We use LLMs to extract knowledge triplets from the given text. See refchecker/extractor for further details and the usage of the knowledge extraction model.

<a id='checking'></a>

Checker

With triplets extracted for the given text, the checker $C$ in the second stage needs to predict hallucination labels on all extracted triplets. We are able to employ numerous existing zero-shot checkers without additional training. We mainly consider two varieties: LLM-based checkers and NLI-based checkers. LLM-based checkers query LLMs to obtain predictions, exhibiting noteworthy success in recent studies. NLI-based checkers adopts much smaller pre-trained language models (such as RoBERTa) to perform text classification, originating from the natural language inference task. See refchecker/checker for more details.

<a id='aggregation'></a>

Aggregation

After collecting fine-grained individual results at triplet-level, we aggregate them to get the overall hallucination label for the entire input text. Based on the definition of the three categories, we define and apply the following intuitive and rigorous aggregation rules in our work.

\tau:Y=\begin{cases} \text{Entailment}&\text{if all } y_i \text{ are Entailment}, \\ \text{Contradiction}&\text{if there exists } y_i \text{ that is Contradiction},\\ \text{Neutral}& \text{otherwise}.\end{cases}

There are many other choices for the aggregation stage. For example, we can count the number of "Neutral" and "Contradiction" triplets, then identify those with the ratio beyond a pre-defined threshold as hallucinated responses. The aggregation rules τ can be customized according to different applications with distinct hallucination tolerance.

<a id='localization'></a>

Evidence Localization

Hallucination detection results can be inferred from certain text spans in the reference, which refers to "evidence". After we obtain the hallucination predictions on triplets, it is meaningful to locate the corresponding evidence in the reference and triplets in the checked text. As explanations for checking results, they provide signals for hallucination mitigation solutions. We use a simple embedding-based method to locate them as certain text spans. It offers an effective baseline for straightforward scenarios where checkers require only a single directly matched evidence for predictions.

<p align="center"><img src="imgs/localization_method.jpg" style="width:300px"></p>

We split reference or the checked text into spans and compare their similarity with elements in triplets. For spans or elements in triplets, we obtain corresponding embeddings as average-pooled token representations by a RoBERTa model. We then compute the pairwise similarity as the cosine distance between a pair of embeddings. Spans matched to elements in triplets with high similarities are considered as localized evidence or triplets.

Localization in Reference

Intuitively, for an entailed claim triplet, we should find all of its constituents in reference. On the other hand, we expect none of them will appear for neutral claim triplets. Finally, contradictory triplets often have missing components. These cases are shown blow:

<p align="center"><img src="imgs/localization_example_1.jpg" style="width:500px"></p> <p align="center"><img src="imgs/localization_example_2.jpg" style="width:500px"></p> <p align="center"><img src="imgs/localization_example_3.jpg" style="width:500px"></p>

Localization Hallucination in Response

Localization in response provides hints for which parts in the response are problematic. We can either conservatively suppress (for "Neutral" triplets) or modify them (for "Contradictory"). Here is an example:

<p align="center"><img src="imgs/localization_example_4.jpg" style="width:500px"></p>

Limitations

Localization is the reverse problem to extraction, and from our experience, making it robust can be a non-trivial research problem. Here are some challenges:

The example below showcases both issues above:

Reference: Bob got his first son Clark in 1982. Two little sisters of Clark were born after 4 years. And he welcomed his youngest son last year.
Triplet: Bob has three children.

In this example, almost the whole reference should be regarded as the evidence. Besides, it requires an implicit addition operation to get the implicit evidence "Bob has four children" that can be directly compared with the triplet.

Limitations and Call for Contributions

Our RefChecker currently has certain limitations that we are actively addressing to enhance the framework. We invite more contributors to join the discussion and share their perspectives on these issues. There are improvements needed for all the components; the RefChecker architecture has a modular architecture that allows independent updates. Here are some helps we felt needed (We welcome contributions through pull requests):

Then there is an array of related problems:

Finally, while RefChecker is in the category of checkers that offer more explainability, it will be very interesting to compare with other approaches that rely on uncertainty at generation time (e.g. SelfCheckGPT).

🔐 Security

See CONTRIBUTING for more information.

🪪 License

The code in this project is licensed under the Apache-2.0 License.