comparia-conversations

<style>
@import url('https://fonts.googleapis.com/css2?family=Marianne:wght@300;400;500;700&display=swap');

:root {
--primary-color: #000091;
--secondary-color: #6a6af4;
--accent-color: #e1000f;
--text-color: #1e1e1e;
--light-bg: #f5f5fe;
--border-radius: 6px;
}

@media (prefers-color-scheme: dark) {
:root {
--primary-color: #6a6af4;
--secondary-color: #8989ff;
--accent-color: #ff5c5c;
--text-color: #e0e0e0;
--light-bg: #252535;
}
}

.container {
font-family: 'Marianne', sans-serif;
max-width: 800px;
margin: 0 auto;
padding: 15px;
color: var(--text-color);
line-height: 1.4;
}

@media (prefers-color-scheme: dark) {
.container {
color: var(--text-color);
}

a {
color: var(--secondary-color);
}

code {
background-color: #333;
color: #f0f0f0;
}
}

.logo {
width: 200px;
display: block;
margin-left: 0;
transition: transform 0.3s ease;
}

.logo:hover {
transform: scale(1.05);
}

h1 {
color: var(--primary-color);
text-align: left;
font-size: 2em;
margin: 40px 0 30px;
position: relative;
padding-bottom: 15px;
}

h1::after {
content: "";
position: absolute;
bottom: 0;
left: 0;
width: 80px;
height: 4px;
background-color: var(--accent-color);
border-radius: 2px;
}

.dataset-section {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(250px, 1fr));
gap: 0px;
margin: 20px 0;
}

.dataset-card {
background-color: white;
border-radius: var(--border-radius);
padding: 25px;
box-shadow: 0 2px 8px rgba(0, 0, 0, 0.05);
display: flex;
flex-direction: column;
}

.stats-cards {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(150px, 1fr));
gap: 20px;
margin: 25px 0;
}

.stat-card {
background-color: white;
border-radius: var(--border-radius);
padding: 20px;
box-shadow: 0 3px 10px rgba(0, 0, 145, 0.1);
text-align: center;
transition: transform 0.2s ease, box-shadow 0.2s ease;
}

.stat-card:hover {
transform: translateY(-3px);
box-shadow: 0 5px 15px rgba(0, 0, 145, 0.15);
}

.stat-card .number {
font-size: 2em;
font-weight: 700;
color: var(--primary-color);
margin: 5px 0;
}

.stat-card .label {
font-size: 0.9em;
color: #555;
font-weight: 500;
}

.dataset-metrics {
background-color: var(--light-bg);
padding: 15px;
border-radius: var(--border-radius);
margin-bottom: 20px;
text-align: center;
display: inline-block;
min-width: 150px;
}

.dataset-metrics .number {
font-size: 1.6em;
font-weight: 700;
color: var(--primary-color);
margin: 4px 0;
}

.dataset-metrics .label {
font-size: 0.9em;
color: #555;
}

.metrics-container {
display: flex;
justify-content: flex-start;
gap: 20px;
flex-wrap: wrap;
margin: 20px 0;
}

.video-container {
box-shadow: 0 4px 12px rgba(0, 0, 145, 0.15);
border-radius: var(--border-radius);
overflow: hidden;
width: 100%;
max-width: 600px;
margin: 20px auto 20px auto;
transition: transform 0.3s ease, box-shadow 0.3s ease;
position: relative;
padding-top: 0;
}

.video-container video {
width: 100%;
display: block;
}

.video-container:hover {
transform: translateY(-3px);
box-shadow: 0 6px 15px rgba(0, 0, 145, 0.2);
}

.highlight-box {
background-color: var(--light-bg);
padding: 20px;
border-radius: var(--border-radius);
margin: 25px 0;
}

.button {
display: inline-block;
background-color: var(--secondary-color);
color: white !important;
text-decoration: none;
padding: 10px 20px;
border-radius: var(--border-radius);
font-weight: 500;
transition: all 0.3s ease;
margin: 5px;
border: 1px solid var(--secondary-color);
}

.button:hover {
transform: translateY(-2px);
box-shadow: 0 4px 8px rgba(0, 0, 0, 0.1);
}

.button.secondary {
background-color: #6A6AF4;
color: white !important;
border: 1px solid var(--primary-color);
}

.datasets-buttons {
display: flex;
justify-content: center;
flex-wrap: wrap;
gap: 10px;
margin: 20px 0;
}

.contact-section {
text-align: left;
margin-top: 40px;
padding: 20px;
background-color: var(--light-bg);
border-radius: var(--border-radius);
}

.data-table {
width: 100%;
border-collapse: collapse;
margin: 20px 0;
}

.data-table th {
background-color: var(--light-bg);
padding: 10px;
text-align: left;
color: var(--primary-color);
}

.data-table td {
padding: 10px;
border-bottom: 1px solid #eee;
}

.datasets-grid {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(300px, 1fr));
gap: 20px;
margin: 15px 0;
}

.datasets-grid .highlight-box {
margin-top: 0;
margin-bottom: 0;
height: 100%;
display: flex;
flex-direction: column;
}

.datasets-grid .highlight-box h3 {
margin-top: 0;
margin-bottom: 10px;
}

.datasets-grid .video-container {
margin: 15px 0 0 0;
max-width: 100%;
}

/* Style pour une vidéo plus élégante et compacte */
.video-container.compact {
max-width: 480px;
margin: 25px auto;
box-shadow: 0 3px 10px rgba(0, 0, 145, 0.1);
border: 1px solid var(--light-bg);
}

.video-container.compact video {
display: block;
width: 100%;
height: auto;
}
</style>
<div class="container">
<a href="https://comparia.beta.gouv.fr/">
<img class="logo" src="https://github.com/user-attachments/assets/bd071ffd-1253-486d-ad18-9f5b371788b0" alt="compar:IA logo">
</a>

# comparia-conversations – dataset containing all questions asked and answers received on the compar:IA platform

## Origin of the data: what is compar:IA?

[Compar:IA](https://comparia.beta.gouv.fr/) is a conversational AI comparison tool (a "chatbot arena") developed within the French Ministry of Culture with a dual mission:
- To educate and raise awareness about the diversity of models, cultural and linguistic biases, and the environmental impact of conversational AIs.
- To improve French-language conversational AIs by publishing French alignment datasets and by creating a ranking of French conversational AI models (in progress).

The Compar:IA platform is developed as part of the State startup compar:IA (incubated by the [Atelier numérique](https://www.culture.gouv.fr/Thematiques/innovation-numerique/Aides-a-l-innovation-et-a-la-transformation-numerique/L-Atelier-numerique#:~:text=L'Atelier%20num%C3%A9rique%20est%20l,engager%20personnellement%20pour%20le%20r%C3%A9soudre.) and [AllIAnce](https://alliance.numerique.gouv.fr/)), integrated into the [beta.gouv.fr](https://beta.gouv.fr) program of the [Interministerial Digital Directorate (DINUM)](https://www.numerique.gouv.fr/dinum/), which helps public administrations build useful, simple, and easy-to-use digital services.
<div style="margin: 20px 0;">
<a href="https://comparia.beta.gouv.fr/" class="button secondary">compar:IA platform website</a>
<a href="https://github.com/betagouv/ComparIA" class="button secondary">compar:IA source code</a>
</div>

## Dataset content

The dataset includes results generated by more than 30 conversational AI models, both open-source and proprietary, of various sizes.
Most questions are asked **in French** and correspond to **real, unconstrained uses** from over 17,345 distinct users.

[](https://app.bunka.ai/datasets/569)

*Source: [Bunka](https://app.bunka.ai/datasets/569).*

## Columns of the comparia-conversations dataset

<table class="data-table">
<tr>
<th>Column</th>
<th>Description</th>
</tr>
<tr>
<td><code>id</code></td>
<td>Unique identifier for each entry in the dataset</td>
</tr>
<tr>
<td><code>timestamp</code></td>
<td>Conversation timestamp</td>
</tr>
<tr>
<td><code>model_a_name</code></td>
<td>Name of the first model</td>
</tr>
<tr>
<td><code>model_b_name</code></td>
<td>Name of the second model</td>
</tr>
<tr>
<td><code>conversation_a</code></td>
<td>Full structure of the conversation with the first model</td>
</tr>
<tr>
<td><code>conversation_b</code></td>
<td>Full structure of the conversation with the second model</td>
</tr>
<tr>
<td><code>conv_turns</code></td>
<td>Number of dialogue turns in the conversation</td>
</tr>
<tr>
<td><code>conversation_pair_id</code></td>
<td>Unique identifier for the pair of conversations</td>
</tr>
<tr>
<td><code>conv_a_id</code></td>
<td>Unique identifier for the conversation with the first model</td>
</tr>
<tr>
<td><code>conv_b_id</code></td>
<td>Unique identifier for the conversation with the second model</td>
</tr>
<tr>
<td><code>session_hash</code></td>
<td>User session identifier</td>
</tr>
<tr>
<td><code>visitor_id</code></td>
<td>Unique anonymized identifier for the user</td>
</tr>
<tr>
<td><code>country</code></td>
<td>Country – empty for now</td>
</tr>
<tr>
<td><code>city</code></td>
<td>City – empty for now</td>
</tr>
<tr>
<td><code>model_pair_name</code></td>
<td>Set representation of the two compared models</td>
</tr>
<tr>
<td><code>opening_msg</code></td>
<td>First message sent by the user</td>
</tr>
<tr>
<td><code>system_prompt_a</code></td>
<td>System prompt for model A</td>
</tr>
<tr>
<td><code>system_prompt_b</code></td>
<td>System prompt for model B</td>
</tr>
<tr>
<td><code>selected_category</code></td>
<td>Prompt suggestion category chosen by the user (if a suggested prompt was used)</td>
</tr>
<tr>
<td><code>is_unedited_prompt</code></td>
<td>Indicates whether the suggested prompt was used as-is</td>
</tr>
<tr>
<td><code>mode</code></td>
<td>Indicates the selected mode</td>
</tr>
<tr>
<td><code>custom_models_selection</code></td>
<td>If the user manually chose models, which ones were selected</td>
</tr>
<tr>
<td><code>short_summary</code></td>
<td>Brief summary of the conversation content</td>
</tr>
<tr>
<td><code>keywords</code></td>
<td>List of keywords summarizing the content</td>
</tr>
<tr>
<td><code>categories</code></td>
<td>Thematic categories of each conversation</td>
</tr>
<tr>
<td><code>languages</code></td>
<td>List of detected languages in the conversation</td>
</tr>
<tr>
<td><code>total_conv_a_output_tokens</code></td>
<td>Total number of output tokens generated by model A during the conversation</td>
</tr>
<tr>
<td><code>total_conv_b_output_tokens</code></td>
<td>Total number of output tokens generated by model B during the conversation</td>
</tr>
<tr>
<td><code>model_a_total_params</code></td>
<td>Estimated or known total size of model A</td>
</tr>
<tr>
<td><code>model_a_acitve_params</code></td>
<td>Estimated or known number of active parameters in model A</td>
</tr>
<tr>
<td><code>model_b_total_params</code></td>
<td>Estimated or known total size of model B</td>
</tr>
<tr>
<td><code>model_b_acitve_params</code></td>
<td>Estimated or known number of active parameters in model B</td>
</tr>
<tr>
<td><code>total_conv_a_kwh</code></td>
<td>Estimated electricity consumption in kWh for model A</td>
</tr>
<tr>
<td><code>total_conv_b_kwh</code></td>
<td>Estimated electricity consumption in kWh for model B</td>
</tr>
</table>

## Purpose of this dataset

We make this dataset available to model developers as well as the research community in artificial intelligence and social sciences to support progress in the following areas:
- Training and alignment of conversational language models, especially in French
- Human–machine interactions and the specific behaviors involved in conversational AI systems
- Improvement of LLM evaluation methods
- AI safety and content moderation

If you use the compar:IA dataset, we would love to hear about your use cases and feedback. It will help us improve the reuse experience.
You can contact us at <a href="mailto:[e-mail]">[e-mail]</a>.

## Data post-processing

User consent is collected through the “Terms of use” section on the website.
A detection of personally identifiable information (PII) was carried out (results are shown in the 'contains_pii' column of the dataset), and conversations containing such information were anonymized.
However, we do not apply any filtering or moderation of potentially toxic or hateful content, to allow researchers to study safety issues related to LLM use in real-world contexts.

## Licenses

Subject to potential third-party claims regarding the language model outputs, the dataset is released under the open Etalab 2.0 license.
It is the responsibility of users to ensure their use of the dataset complies with applicable laws and regulations, especially regarding data protection and the terms of use of different model providers.

## Other compar:IA datasets

<div class="datasets-grid">
<div class="highlight-box">
<h3>comparIA-reactions</h3>
<p>Dataset collecting user reactions to compar:IA at the message level. It reflects preferences expressed throughout conversations, message by message.</p>
<div class="video-container">
<video controls autoplay loop muted playsinline src="https://cdn-uploads.huggingface.co/production/uploads/649d986a474bf415c03b772c/ncldPIO_bTesSd8bqcjqn.mp4"></video>
</div>
<div class="datasets-buttons">
<a href="https://huggingface.co/datasets/ministere-culture/comparia-reactions" class="button secondary">Explore comparIA-reactions</a>
</div>
</div>

<div class="highlight-box">
<h3>comparIA-votes</h3>
<p>Dataset grouping user preferences expressed after a full conversation. These votes compare two models when no individual message reactions are given.</p>
<div class="video-container">
<video controls autoplay loop muted playsinline src="https://cdn-uploads.huggingface.co/production/uploads/649d986a474bf415c03b772c/SQaFiHnt8LPx0rxwl6vVG.mp4"></video>
</div>
<div class="datasets-buttons">
<a href="https://huggingface.co/datasets/ministere-culture/comparia-votes" class="button secondary">Explore comparIA-votes</a>
</div>
</div>
</div>

<div class="contact-section">
<h3>Reporting sensitive data</h3>
<p>If you find a line in the dataset that you think contains PII or sensitive data, please let us know via <a href="https://adtk8x51mbw.eu.typeform.com/to/B49aloXZ">this short form</a>.</p>

<h3>Contact</h3>
<p>For any questions or requests for information, contact <a href="mailto:[e-mail]">[e-mail]</a></p>

<div style="margin-top: 30px;">
<a href="https://beta.gouv.fr">
<img src="https://cdn-uploads.huggingface.co/production/uploads/649d986a474bf415c03b772c/Zk4YiqgKu9sm5ydQ7fhSq.png" alt="Logo of the Ministry, beta.gouv and Atelier numérique" style="max-width: 400px;">
</a>
</div>
</div>

</div>

<div align="center">

<br />
<a href="https://digitalpublicgoods.net/r/comparia" target="_blank" rel="noopener noreferrer"><img src="https://github.com/DPGAlliance/dpg-resources/blob/main/docs/assets/dpg-badge.png?raw=true" width="100" alt="Digital Public Goods Badge"></a>

</div>

# comparia-conversations – 收录compar:IA平台所有用户提问与平台回复的数据集

## 数据来源：compar:IA平台简介
[Compar:IA](https://comparia.beta.gouv.fr/) 是一款对话式AI（conversational AI）对比工具（又称“聊天机器人竞技场”），由法国文化部开发，肩负双重使命：
- 科普对话式AI的模型多样性、文化与语言偏见，以及其环境影响，提升公众认知；
- 通过发布法语对齐数据集，以及构建法语对话式AI模型排名（开发中），优化法语对话式AI生态。

Compar:IA平台作为法国国家初创项目compar:IA的一部分开发，由[数字工坊(Atelier numérique)](https://www.culture.gouv.fr/Thematiques/innovation-numerique/Aides-a-l-innovation-et-a-la-transformation-numerique/L-Atelier-numerique#:~:text=L'Atelier%20num%C3%A9rique%20est%20l,engager%20personnellement%20pour%20le%20r%C3%A9soudre.)与[AllIAnce](https://alliance.numerique.gouv.fr/)孵化，隶属于跨部委数字总局（Interministerial Digital Directorate, DINUM）的[beta.gouv.fr](https://beta.gouv.fr)计划。该计划旨在帮助公共行政部门打造实用、简洁且易用的数字化服务。

<div style="margin: 20px 0;">
<a href="https://comparia.beta.gouv.fr/" class="button secondary">compar:IA平台官网</a>
<a href="https://github.com/betagouv/ComparIA" class="button secondary">compar:IA源代码</a>
</div>

## 数据集内容
本数据集涵盖超过30款不同规模的开源与专有对话式AI模型生成的对话结果。绝大多数提问以法语发起，均来自17345余名独立用户的真实、无约束使用场景。

[](https://app.bunka.ai/datasets/569)
*来源：[Bunka](https://app.bunka.ai/datasets/569)。*

## comparia-conversations数据集字段说明
<table class="data-table">
<tr>
<th>字段名</th>
<th>描述</th>
</tr>
<tr>
<td><code>id</code></td>
<td>数据集内每条数据的唯一标识符</td>
</tr>
<tr>
<td><code>timestamp</code></td>
<td>对话时间戳</td>
</tr>
<tr>
<td><code>model_a_name</code></td>
<td>首个对比模型名称</td>
</tr>
<tr>
<td><code>model_b_name</code></td>
<td>第二个对比模型名称</td>
</tr>
<tr>
<td><code>conversation_a</code></td>
<td>与首个模型的完整对话结构</td>
</tr>
<tr>
<td><code>conversation_b</code></td>
<td>与第二个模型的完整对话结构</td>
</tr>
<tr>
<td><code>conv_turns</code></td>
<td>对话轮次总数</td>
</tr>
<tr>
<td><code>conversation_pair_id</code></td>
<td>对话对的唯一标识符</td>
</tr>
<tr>
<td><code>conv_a_id</code></td>
<td>与首个模型的对话唯一标识符</td>
</tr>
<tr>
<td><code>conv_b_id</code></td>
<td>与第二个模型的对话唯一标识符</td>
</tr>
<tr>
<td><code>session_hash</code></td>
<td>用户会话标识符</td>
</tr>
<tr>
<td><code>visitor_id</code></td>
<td>用户的匿名化唯一标识符</td>
</tr>
<tr>
<td><code>country</code></td>
<td>国家信息（当前为空）</td>
</tr>
<tr>
<td><code>city</code></td>
<td>城市信息（当前为空）</td>
</tr>
<tr>
<td><code>model_pair_name</code></td>
<td>两个对比模型的集合表示</td>
</tr>
<tr>
<td><code>opening_msg</code></td>
<td>用户发送的首条消息</td>
</tr>
<tr>
<td><code>system_prompt_a</code></td>
<td>模型A的系统提示词</td>
</tr>
<tr>
<td><code>system_prompt_b</code></td>
<td>模型B的系统提示词</td>
</tr>
<tr>
<td><code>selected_category</code></td>
<td>用户选择的提示词建议类别（若使用了推荐提示词）</td>
</tr>
<tr>
<td><code>is_unedited_prompt</code></td>
<td>指示是否直接使用了推荐的提示词</td>
</tr>
<tr>
<td><code>mode</code></td>
<td>指示所选交互模式</td>
</tr>
<tr>
<td><code>custom_models_selection</code></td>
<td>若用户手动选择模型，则记录所选模型列表</td>
</tr>
<tr>
<td><code>short_summary</code></td>
<td>对话内容的简要摘要</td>
</tr>
<tr>
<td><code>keywords</code></td>
<td>总结对话内容的关键词列表</td>
</tr>
<tr>
<td><code>categories</code></td>
<td>对话所属的主题类别</td>
</tr>
<tr>
<td><code>languages</code></td>
<td>对话中检测到的语言列表</td>
</tr>
<tr>
<td><code>total_conv_a_output_tokens</code></td>
<td>模型A在对话中生成的输出Token总数</td>
</tr>
<tr>
<td><code>total_conv_b_output_tokens</code></td>
<td>模型B在对话中生成的输出Token总数</td>
</tr>
<tr>
<td><code>model_a_total_params</code></td>
<td>模型A的总参数量（估算值或已知值）</td>
</tr>
<tr>
<td><code>model_a_active_params</code></td>
<td>模型A的有效参数量（估算值或已知值，原文原笔误acitve已修正）</td>
</tr>
<tr>
<td><code>model_b_total_params</code></td>
<td>模型B的总参数量（估算值或已知值）</td>
</tr>
<tr>
<td><code>model_b_active_params</code></td>
<td>模型B的有效参数量（估算值或已知值，原文原笔误acitve已修正）</td>
</tr>
<tr>
<td><code>total_conv_a_kwh</code></td>
<td>模型A在对话中产生的估算电力消耗量（单位：千瓦时）</td>
</tr>
<tr>
<td><code>total_conv_b_kwh</code></td>
<td>模型B在对话中产生的估算电力消耗量（单位：千瓦时）</td>
</tr>
</table>

## 数据集用途
本数据集面向模型开发者以及人工智能与社会科学领域的研究社区开放，以支持以下方向的研究进展：
- 对话式语言模型的训练与对齐，尤其是法语方向；
- 人机交互及对话式AI系统特有的行为模式研究；
- 大语言模型（Large Language Model, LLM）评估方法的优化；
- AI安全与内容审核。

若您使用compar:IA数据集，我们期待了解您的应用场景与反馈，这将帮助我们优化数据集的复用体验。您可通过邮箱<a href="mailto:[e-mail]">[e-mail]</a>联系我们。

## 数据后处理
用户同意通过平台的“使用条款”页面收集其数据。我们已执行个人身份信息（Personally Identifiable Information, PII）检测，检测结果将在数据集的`contains_pii`字段中体现，并对包含此类信息的对话进行了匿名化处理。但我们未对潜在的有毒或仇恨性内容进行过滤或审核，以确保研究人员能够真实研究大语言模型应用中的安全问题。

## 许可协议
除可能存在的第三方针对模型输出的主张外，本数据集采用开放的Etalab 2.0协议发布。用户需自行确保数据集的使用符合适用法律法规，尤其需遵守数据保护相关要求以及各模型服务商的使用条款。

## 其他compar:IA数据集
<div class="datasets-grid">
<div class="highlight-box">
<h3>comparIA-reactions</h3>
<p>本数据集收录用户在消息层面针对compar:IA平台的反馈，反映用户在对话过程中针对单条消息表达的偏好。</p>
<div class="video-container">
<video controls autoplay loop muted playsinline src="https://cdn-uploads.huggingface.co/production/uploads/649d986a474bf415c03b772c/ncldPIO_bTesSd8bqcjqn.mp4"></video>
</div>
<div class="datasets-buttons">
<a href="https://huggingface.co/datasets/ministere-culture/comparia-reactions" class="button secondary">探索comparIA-reactions</a>
</div>
</div>

<div class="highlight-box">
<h3>comparIA-votes</h3>
<p>本数据集收录用户在完整对话结束后表达的偏好，此类投票用于对比两个模型，且不包含单条消息的反馈。</p>
<div class="video-container">
<video controls autoplay loop muted playsinline src="https://cdn-uploads.huggingface.co/production/uploads/649d986a474bf415c03b772c/SQaFiHnt8LPx0rxwl6vVG.mp4"></video>
</div>
<div class="datasets-buttons">
<a href="https://huggingface.co/datasets/ministere-culture/comparia-votes" class="button secondary">探索comparIA-votes</a>
</div>
</div>
</div>

<div class="contact-section">
<h3>敏感数据上报</h3>
<p>如果您在数据集中发现包含个人身份信息或敏感数据的条目，请通过<a href="https://adtk8x51mbw.eu.typeform.com/to/B49aloXZ">此简短表单</a>告知我们。</p>

<h3>联系方式</h3>
<p>如有任何问题或信息请求，请联系<a href="mailto:[e-mail]">[e-mail]</a></p>

<div style="margin-top: 30px;">
<a href="https://beta.gouv.fr">
<img src="https://cdn-uploads.huggingface.co/production/uploads/649d986a474bf415c03b772c/Zk4YiqgKu9sm5ydQ7fhSq.png" alt="法国文化部、beta.gouv与数字工坊联合标识" style="max-width: 400px;">
</a>
</div>
</div>

<div align="center">
<br />
<a href="https://digitalpublicgoods.net/r/comparia" target="_blank" rel="noopener noreferrer"><img src="https://github.com/DPGAlliance/dpg-resources/blob/main/docs/assets/dpg-badge.png?raw=true" width="100" alt="数字公共产品徽章"></a>
</div>