language:

• fr
• en
• es
  license: cc-by-4.0
  size_categories:
• 1K<n<10K
  task_categories:
• text-generation
  pretty_name: Phare
  configs:
• config_name: hallucination_tools_basic
  data_files:
  • split: public
    path: hallucination/tools/basic.parquet
• config_name: hallucination_tools_knowledge
  data_files:
  • split: public
    path: hallucination/tools/knowledge.parquet
• config_name: hallucination_debunking
  data_files:
  • split: public
    path: hallucination/debunking/*.parquet
• config_name: hallucination_factuality
  data_files:
  • split: public
    path: hallucination/factuality/*.parquet
• config_name: hallucination_satirical
  data_files:
  • split: public
    path: hallucination/satirical/*.parquet
• config_name: harmful_vulnerable_misguidance
  data_files:
  • split: public
    path: harmful/vulnerable_misguidance/*.parquet
• config_name: biases
  data_files:
  • split: public
    path: biases/story_generation/*.parquet
• config_name: jailbreak_encoding
  data_files:
  • split: public
    path: jailbreak/jailbreak/encoding.parquet
• config_name: jailbreak_framing
  data_files:
  • split: public
    path: jailbreak/jailbreak/framing.parquet
• config_name: jailbreak_injection
  data_files:
  • split: public
    path: jailbreak/jailbreak/injection.parquet

Phare Benchmark

Phare is a multilingual benchmark that measures LLM Safety across multiple categories of vulnerabilities, including hallucination, biases & stereotypes, harmful content, and jailbreaks.

Dataset Details

Dataset Description

This dataset contains the public set of samples of Phare Benchmark. These samples are split into multiple modules to assess LLM safety across various directions.
Each module is responsible for detecting vulnerabilities in the LLM response:

• Hallucination: evaluates the factuality and the level of misinformation spread by the models in a question-answer setting. Questions are designed from existing content, including known misinformation or scientifically refuted theories.

• Biases & stereotypes: assesses the presence of biases in the LLM generations for creative tasks.

• Harmful content: measures the dangerous behavior endorsement and misguidance rate of LLM with vulnerable people.

• Jailbreak: measures the robustness of the LLMs against various types of jailbreak attacks (e.g, encoding, role-play, known injection, etc.).

Each module is split into several submodules. The submodules are different approaches to eliciting problematic behavior from the models. For instance, the hallucination modules has several submodules:

• Debunking: questions about scientifically refuted facts or theories with various levels of bias
• Satirical: questions derived from misinformation and satirical sources
• Factuality: questions about generic facts
• Tools: questions that can be answered with the use of a tool available for the model, to measure hallucination in tool parameters and correct tool usage.

Extra information

• Author: Giskard AI
• Language(s): English, French, Spanish
• License: CC BY 4.0, except for the Jailbreak module, which is licensed under AGPL-3.0

Dataset Structure

The dataset is split into a public (available in this repository) and a private sets. Giskard reserves the private set to run the Phare Benchmark and keep the leaderboard up-to-date.

Each submodule is a set of .jsonl files containing the samples.
Each sample in these files has the following structure:

{
  "id": # unique id of the sample
  "module": # the module the sample belongs to
  "task": # the submodule or task the sample belongs to
  "language": # the language of the sampel
  "generations": # a list of generations/completions that will be performed for the given sample
      [{
        "id": # unique id of the generation
        "type": # the type of completion, currently only support s`chat_completion`
        "messages": # the list of messages to complete
        "params": # a dictionary with optional generation params (temperature, tools, etc.)
  }]
  "metadata": # a dictionary of metadata, with mainly information about the generation process of the sample
  "evaluation": # a dictionary with additional elements required for the automatic evaluation of the response (include context about the question, expected answers, etc.), changes from between submodules.
}

Dataset Creation

Curation Rationale

Most safety evaluation datasets lack comprehensiveness and multicultural support. Our main goal with Phare is to fill this gap and propose a benchmark that detects inappropriate behavior in various situations.
In addition, the dataset was designed in multiple languages from scratch, including the data collection phase to ensure multicultural diversity.

Source Data

Data sources are diverse and change for each module:

• Hallucinations: news articles, wikipedia articles, satirical articles, forum threads, etc.
• Harmful content: examples of AI incident from https://incidentdatabase.ai/
• Biases & Stereotypes: legal documents about discriminatory attributes.
• Jailbreak: existing dataset of well known attacks, precise attributions are available under jailbreak/LICENSE.md.

The Hallucination module uses the source data more extensively than other modules. The hallucination questions are grounded on existing content, while for other modules, the data source only influences the evaluation process, e.g. legislation about discrimination fixes the attributes that are extracted from the LLM answers.

Data Collection and Processing

Data collection and filtering were done semi-automatically by the Giskard team. The initial steps of data collection and filtering were done automatically with diverse criteria depending on the module.

Following the data collection and filtering step, the samples are generated using diverse strategies. It includes a combination of LLM generation and the application of handcrafted templates. All details about the generation process are available in our technical report.

A manual review was then conducted on the generated samples by native speakers of the corresponding language to make sure the samples were meeting our quality criteria.

Personal and Sensitive Information

The dataset contains samples that can be sensitive or misleading. In particular, the harmful content module contains samples that push for dangerous behavior. Similarly, the hallucination module contains samples made of factually false content.

Bias, Risks, and Limitations

• Some content was generated with the help of LLMs and can be imperfect.
• Some data sources used in particular in the hallucination module can be partial.
• The evaluation process is automatic and is not fully accurate, we measured the evaluation quality manually on each submodule individually to ensure the errors are constrained. Results of this analysis will be reported precisely in the technical paper.
• The team that manually reviewed the samples and the evaluation results has limited representativity.
• Some modules and languages have more samples than others and will have more influence on the aggregated scores.
• Private and public splits representativity differs across modules.

Dataset Card Contact

• Matteo Dora -- @mattbit -- [email protected]
• Pierre Le Jeune -- @pierlj -- [email protected]