Skip to content

Latest commit

 

History

History
302 lines (237 loc) · 16.5 KB

README.md

File metadata and controls

302 lines (237 loc) · 16.5 KB
Raw

NoMIRACL: A Multilingual Hallucination Evaluation Dataset for Robust RAGs

Stars Build License arXiv


The image has been generated using miramuseai.net and Adobe Photoshop.

NoMIRACL is multilingual hallucination evaluation dataset across 18 diverse languages. It includes both a non-relevant and a relevant subset. The non-relevant subset contains queries with passages manually judged as non-relevant, while the relevant subset includes queries with at least one judged relevant passage. LLM robustness is measured using two key metrics: hallucination rate and error rate.

This repository provides easy code to implement and evaluate diverse LLM baselines using our prompt template on NoMIRACL.

For more information, checkout out our publication:

🔧 Installation

You can install NoMIRACL via pip:

pip install nomiracl

If you want to build from source, use:

$ git clone https://github.com/project-miracl/nomiracl.git
$ cd nomiracl
$ pip install -e .

⭐ Getting Started

1. Load NoMIRACL Dataset

  • 50% of relevant examples, 50% of non-relevant, both maximum capped at 250.
  • Full example available in sample_load_no_miracl.py.
from nomiracl.dataset import NoMIRACLDataLoader

data_loader = NoMIRACLDataLoader(language = "english, 
                                 split = "test", # or 'dev' 
                                 hf_dataset_name="miracl/nomiracl", 
                                 load_from_huggingface=True)
                        
corpus, queries, qrels = data_loader.load_data_sample(
    relevant_ratio = 0.5, non_relevant_ratio = 0.5, max_sample_pool = 250)

2. LLM prompt generation

from nomiracl.generation.utils import load_model

model_name = "zephyr-7b-beta"
weights_path = f"HuggingFaceH4/{model_name}"
model = load_model(model_name, weights_path=weights_path, cache_dir=None)

# Sample prompts
prompts = [
    "What is the capital of France?",
    "What is the capital of Germany?",
    "What is the capital of Italy?",
]

model_results = model.batch_call(prompts, batch_size=1)

for prompt, result in zip(prompts, model_results):
    print("Prompt: {}".format(prompt))
    print("{} result: {}".format(model_name, result))

3. Loading Vanilla prompt template

from nomiracl.prompts.utils import load_prompt_template

prompt_cls = load_prompt_template("vanilla", count = 10)

query = "Which is the best programming language?"

passages = [
    "Python is the best programming language.",
    "Javascript is the best programming language.",
    "Go is the best programming language.",
    "Java is the best programming language.",
    "C# is the best programming language.",
    "Ruby is the best programming language.",
    "R is the best programming language.",
    "C++ is the best programming language.",
    "C is the best programming language.",
    "Rust is the best programming language.",
]

prompt = prompt_cls(query=query, passages=passages)
print(prompt)

🤗 NoMIRACL Dataset

The NoMIRACL dataset is available in HuggingFace under: miracl/nomiracl.

Languages covered: Arabic (ar), Bengali (bn), German (de), English (en), Spanish (es), Persian (fa), Finnish (fi), French (fr), Hindi (hi), Indonesian (id), Japanese (ja), Korean (ko), Russian (ru), Swahili (sw), Thai (th), Yoruba (yo), Chinese (zh).

HuggingFace Page: https://huggingface.co/datasets/miracl/nomiracl

import datasets

language = 'german'  # or any of the 18 languages
subset = 'relevant'  # or 'non_relevant'
split = 'test'       # or 'dev' for development split

# four combinations available: 'dev.relevant', 'dev.non_relevant', 'test.relevant' and 'test.non_relevant'
nomiracl = datasets.load_dataset('miracl/nomiracl', language, split=f'{split}.{subset}')

Model identifiers for evaluation in NoMIRACL

Acronym Model Name Model Link
GPT-4 gpt-4-azure AzureAI
GPT-3.5 gpt-3.5-azure AzureAI
Mixtral-7x8B Mixtral-8x7B-Instruct-v0.1 🤗 model
Mistral-7B Mistral-7B-Instruct-v0.2 🤗 model
Orca-2-13B Orca-2-13b 🤗 model
Orca-2-7B Orca-2-7b 🤗 model
Aya-101 aya-101 🤗 model
LLAMA-2-70B llama-2-70b-chat 🤗 model
LLAMA-2-13B llama-2-13b-chat 🤗 model
LLAMA-2-7B llama-2-7b-chat 🤗 model
Flan-T5-XXL flan-t5-xxl 🤗 model

Baseline Accuracy on NoMIRACL non-relevant subset (test split, maximum cap of 250 per language)

Baseline results (250 queries) are available within the repository under ./results/baselines/non_relevant.

An example datapoint under ./results/baselines/non_relevant/en.test.vanilla_prompt.jsonl

{
    "query_id": "842558#0", 
    "docids": ["2842207#5", "7004944#45", "3310762#14", "47220460#1", "36451733#7", "3310762#20", "4724576#4", "22373402#0", "52203230#0", "23126218#4"], 
    "prompt": "I will give you a question and several contexts containing information about the question. [ ... ] \n\nOUTPUT:\n", 
    "template": "vanilla", 
    "results": {"gpt-4-azure": "Yes, answer is present.", 
                "llama-2-13b-chat": "\nYes, answer is present in [6].\n\nNo answers found in the other contexts.",
                [...]
                "aya-101": "Wales"}
}
Code Set #Q GPT-4 GPT-3.5 Mistral-7B Mixtral-7x8B LLAMA-2-70B LLAMA-2-13B LLAMA-2-7B flan-t5-xxl Orca-2-7B Orca-2-13B aya-101
ar Test 250 61.60% 46.40% 87.20% 89.20% 0.0% 15.60% 0.0% 0.80% 3.60% 10.40% 23.20%
bn Test 250 60.00% 4.80% 83.20% 90.00% 0.0% 2.40% 0.4% 0.00% 5.60% 3.20% 10.00%
de Test 217 63.59% 53.00% 87.56% 68.20% 0.5% 5.07% 0.9% 3.23% 5.07% 12.90% 29.03%
en Test 250 57.20% 54.80% 90.00% 72.40% 0.0% 0.80% 2.8% 16.40% 12.00% 6.80% 15.60%
es Test 250 87.20% 64.80% 92.00% 90.80% 0.8% 0.40% 11.2% 10.80% 14.40% 10.40% 3.20%
fa Test 250 57.20% 23.60% 82.40% 90.40% 0.0% 4.80% 0.0% 0.40% 0.40% 14.00% 14.40%
fr Test 250 52.40% 44.00% 82.40% 58.40% 0.0% 0.00% 0.4% 2.40% 6.00% 9.20% 22.00%
fi Test 124 60.48% 65.32% 87.90% 89.52% 0.0% 4.84% 0.0% 0.00% 2.42% 27.42% 33.06%
hi Test 250 78.80% 29.60% 91.60% 95.60% 0.0% 3.20% 0.8% 0.00% 0.40% 9.20% 17.60%
id Test 250 63.20% 56.80% 89.20% 83.20% 0.4% 4.80% 1.6% 6.80% 2.80% 14.40% 19.60%
ja Test 250 56.80% 32.40% 89.20% 82.80% 0.0% 4.00% 0.0% 0.80% 7.60% 24.00% 10.40%
ko Test 250 59.60% 40.00% 88.40% 90.00% 0.0% 0.80% 1.2% 0.00% 3.60% 10.80% 14.40%
ru Test 250 58.00% 34.80% 90.00% 78.40% 0.8% 4.00% 0.4% 1.60% 11.20% 9.20% 31.60%
sw Test 250 91.20% 66.40% 95.20% 88.00% 0.0% 0.80% 0.4% 7.60% 4.40% 14.00% 27.60%
te Test 250 74.80% 6.80% 81.20% 84.80% 0.0% 0.40% 0.0% 1.60% 6.80% 8.00% 24.00%
th Test 250 46.80% 4.00% 90.40% 67.20% 0.0% 16.40% 0.0% 0.80% 4.40% 5.60% 11.20%
yo Test 250 75.20% 74.80% 95.20% 89.20% 0.0% 1.20% 0.4% 12.80% 13.60% 14.80% 20.00%
zh Test 250 56.40% 43.60% 86.80% 78.40% 0.0% 6.00% 0.0% 5.20% 4.40% 10.80% 9.20%
Avg. Test - 64.5% 41.44% 88.33% 82.6% 0.1% 4.2% 1.14% 3.96% 6.04% 11.95% 18.67%

Baseline Accuracy on NoMIRACL relevant subset (test split, maximum cap of 250 per language)

Baseline results (250 queries) are available within the repository under ./results/baselines/relevant.

An example datapoint under ./results/baselines/relevant/en.test.vanilla_prompt.jsonl

{
    "query_id": "8706103#0", 
    "docids": ["42057469#2", "4998067#1", "29247933#0", "162619#81", "422315#13", "26790310#4", "41298602#18", "22816#16", "123427#61", "23576525#0"], 
    "prompt": "I will give you a question and several contexts containing information about the question. [ ... ] \n\nQUESTION:\nWhat is the course that will be discontinued as defined by the National Education Policy? [ ... ] \n\nOUTPUT:\n", 
    "template": "vanilla", 
    "results": {"gpt-4-azure": "I don't know.", 
                "llama-2-13b-chat": "Please answer the question based on the given contexts.",
                [...]
                "aya-101": "I don't know"}
}
Code Set #Q GPT-4 GPT-3.5 Mistral-7B-v0.2 Mixtral-7x8B LLAMA-2-70B LLAMA-2-13B LLAMA-2-7B flan-t5-xxl Orca-2-7B Orca-2-13B aya-101
ar Test 250 88.40% 91.20% 32.80% 59.20% 96.40% 62.40% 62.0% 100.0% 82.80% 51.60% 16.40%
bn Test 250 82.80% 64.80% 26.40% 43.20% 97.60% 42.40% 71.6% 100.0% 40.00% 80.00% 20.00%
de Test 217 88.40% 93.60% 26.80% 74.40% 88.80% 52.80% 56.4% 99.2% 86.00% 92.00% 25.60%
en Test 250 94.80% 91.20% 35.20% 78.80% 85.20% 46.40% 68.0% 98.8% 81.60% 98.00% 51.60%
es Test 250 77.60% 90.00% 26.80% 67.20% 77.20% 34.40% 48.8% 99.2% 64.00% 95.60% 78.00%
fa Test 250 86.40% 95.60% 30.80% 46.80% 99.20% 85.60% 72.0% 100.0% 84.40% 61.60% 24.40%
fr Test 250 88.40% 90.40% 37.60% 81.60% 92.00% 54.80% 55.6% 99.6% 81.60% 98.00% 57.20%
fi Test 124 84.00% 87.60% 24.80% 65.20% 98.40% 6.00% 93.2% 100.0% 79.20% 75.20% 23.60%
hi Test 250 78.80% 92.80% 28.00% 38.40% 94.80% 66.00% 74.8% 100.0% 57.60% 57.60% 25.20%
id Test 250 66.00% 74.00% 11.20% 48.00% 83.60% 32.00% 71.2% 100.0% 73.20% 79.20% 60.71%
ja Test 250 95.60% 97.20% 31.20% 69.20% 98.40% 52.40% 69.2% 100.0% 62.80% 64.80% 45.60%
ko Test 250 87.20% 92.40% 22.40% 56.00% 99.60% 90.00% 85.6% 100.0% 77.60% 80.00% 60.00%
ru Test 250 93.60% 94.40% 32.40% 77.20% 83.20% 61.60% 96.4% 100.0% 79.60% 85.20% 22.80%
sw Test 250 78.80% 90.40% 8.40% 51.60% 90.00% 62.80% 49.2% 100.0% 89.20% 92.80% 38.40%
te Test 250 58.00% 45.60% 14.80% 33.20% 99.60% 74.40% 97.2% 100.0% 44.80% 73.60% 10.80%
th Test 250 95.60% 96.40% 23.60% 72.80% 98.80% 59.20% 91.2% 100.0% 78.00% 76.00% 20.80%
yo Test 250 85.78% 64.22% 8.33% 34.31% 91.18% 62.25% 62.3% 99.5% 79.41% 82.84% 18.14%
zh Test 250 95.60% 95.60% 30.00% 74.00% 97.20% 63.20% 83.6% 100.0% 81.20% 90.40% 75.20%
Avg. Test - 84.77% 85.97% 25.09% 59.51% 92.84% 56.04% 72.68% 99.8% 73.50% 79.69% 37.47%

NoMIRACL Dataset Construction

Retrieval Augmented Generation (RAG) is a powerful approach to incorporate external knowledge into large language models (LLMs) to enhance the accuracy and faithfulness of generated responses. However, evaluating LLM robustness in RAG across different language families has been a challenge, leading to gaps in understanding the model's performance against errors in external retrieved knowledge. To address this, we present NoMIRACL, a human-annotated dataset designed for evaluating LLM robustness in RAG across 18 typologically diverse languages.

NoMIRACL is a multilingual dataset designed to evaluate LLM robustness against errors in first-stage retrieval. The dataset covers 18 typologically diverse languages and includes two subsets: non-relevant and relevant.

Non-Relevant Subset (F)

  • Queries with no-known answers.
  • All top-k passages manually judged as non-relevant (relevancy score = 0).

Relevant Subset (T)

  • Queries with known answers.
  • At least one of the top-k passages manually judged as relevant (relevancy score = 1).

Evaluation Metrics

We conduct a robustness evaluation using a binary classification task, comparing LLM predictions against the ground truth provided in NoMIRACL. The metrics used are hallucination rate and error rate.

  • Hallucination Rate: FP/(FP + TN) Measures the model's tendency to hallucinate an answer when no answer is present in the non-relevant subset.

  • Error Rate: FN/(FN + TP) Measures the model's inaccuracy in recognizing relevant passages in the relevant subset.

🤝 Collaboration and Acknowledgements

The NoMIRACL dataset has been made possible due to a collaborative effort of the following universities and organizations:

  • University of Waterloo
  • Huawei Noah's Ark Lab

Parts of the NoMIRACL code structure has been inspired by:

📜 Citations

If you use NoMIRACL or parts in a research paper, please cite our work as follows:

@article{thakur:2024,
  author       = {Nandan Thakur and
                  Luiz Bonifacio and
                  Xinyu Zhang and
                  Odunayo Ogundepo and
                  Ehsan Kamalloo and
                  David Alfonso{-}Hermelo and
                  Xiaoguang Li and
                  Qun Liu and
                  Boxing Chen and
                  Mehdi Rezagholizadeh and
                  Jimmy Lin},
  title        = {NoMIRACL: Knowing When You Don't Know for Robust Multilingual Retrieval-Augmented
                  Generation},
  journal      = {CoRR},
  volume       = {abs/2312.11361},
  year         = {2023},
  url          = {https://doi.org/10.48550/arXiv.2312.11361},
  doi          = {10.48550/ARXIV.2312.11361},
  eprinttype    = {arXiv},
  eprint       = {2312.11361},
  timestamp    = {Tue, 16 Jan 2024 11:57:42 +0100},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2312-11361.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

Contact person: Nandan Thakur, nandan.thakur@uwaterloo.co

This repository contains experimental software and is published for the sole purpose of giving additional background details on the respective publication.