--- language: - fi license: other size_categories: - n<1K task_categories: - conversational dataset_info: features: - name: System dtype: string - name: Prompt dtype: string - name: Replies [1] dtype: string - name: Replies [2] dtype: string - name: Replies [3] dtype: string - name: Replies [4] dtype: string splits: - name: train num_bytes: 162473 num_examples: 366 download_size: 77390 dataset_size: 162473 license_name: tuumailubotti-licence licence-link: https://huggingface.co./datasets/Yleisradio/Tuumailubotti/blob/main/LICENCE.md configs: - config_name: default data_files: - split: train path: data/train-* --- ## Dataset Description - **Point of Contact:**  johan.sundstrom@yle.fi, vertti.luostarinen@gmail.com ### Dataset Summary Tuumailubotti dataset is a Finnish-speaking conversational dataset containing 364 rows of simulated reflective conversations about work life from a neurodiversity-affirming perspective. Each row of the dataset starts with a system prompt, followed by either four or five exchanges between bot and user. The dataset was commissioned as part of the Tuumailubotti project by the Finnish Broadcasting Company, YLE, and features data both from contributors from inside and outside the company. ### Supported Tasks - `conversational`: The dataset can be used to train a model for Conversational tasks, such as chatbots. We trained it for this purpose using TurkuNLP/gpt3-finnish-8B (with 8-bit quantization). https://huggingface.co./TurkuNLP/gpt3-finnish-8B For the most part, the dataset can also be utilized in a question answering context, but please note that compared to traditional q&a datasets the roles are reversed. ### Languages - `fi` The dataset has a mix of colloquial and formal Finnish. ## Dataset Structure ### Data Instances ``` { 'System': ..., 'Bot1': ..., 'User1': ..., 'Bot2': ..., 'User2':..., 'Bot3':... ... } ``` We collected the data originally in the OASST format with multiple responses. We converted this structure to a tabular format for training, as at this phase we were not interested in reinforcement learning. If the project goes forward, we might release a new version of the dataset with a branching structure. ### Data Fields - `System`: String, The system prompt for the current conversation; a very brief description of the purpose of the bot and its task. - `Bot1`: String, First turn of dialogue. - `User1`: String, Second turn of dialogue. - `Bot2`: String, Third turn of dialogue. - `User2`: String, Fourth turn of dialogue. - `Bot3`: String, Fifth turn of dialogue, sometimes left empty. There are no splits in the dataset. ## Dataset Creation ### Curation Rationale This dataset contains simulated conversations about workplace wellbeing and work life in general. It has been assembled for the purpose of creating a neurodiversity-affirming workplace reflection chatbot that can help employees to contemplate questions they could never ask from a human HR representative. ### Source Data The data contributors were instructed to imagine what kind of conversations they personally would like to have with the finished bot, and then write down these fictitious conversations without mentioning any real-world personal data. We provided three topics, based on the "TAOS" development discussion framework: goals, wellbeing and know-how. Additionally, the contributors were also encouraged to come up with their own topics. Because of this, the data also includes more broad discussion about neurodiversity and the impact of today's competitive society on individual lives. As we found out that the base Finnish GPT models contain either outdated or outright harmful information about neurodiversity-related topics, we also included some information that utilized the vocabulary created by Autistic Spectrum Finland (ASY). #### Initial Data Collection and Normalization While obvious typos have been fixed from the responses of the bot and the system prompts, the dataset has not otherwise been filtered or altered. This phase was omitted on purpose in order to research the potential diversity ingrained in this dataset. #### Who are the source language producers? The data was provided by contributors both inside and outside of YLE, with a roughly 70-30-split. The data was gathered from anonymous employees from YLE's neurodiversity network, as well as known persons affiliated with the project. The outside contributors all identify as members of neurominorities and their identities are known, though they wish to remain anonymous. All major contributors (those who contributed more than one or two samples ) were compensated for their time. ### Personal and Sensitive Information Because of the delicate subject matter, no diagnostic information, personal information etc. was gathered about the contributors. While we know that many contributors either identify as members of neurominorities and/or have i.e.. an autism or ADHD diagnosis, no identifying information was featured in the data. This was also done because we forbid this data to be used in medical research or in other ways in which it could be used to otherwise identify or segregate members of neurominorities. The contributors were instructed to pretend to be fictitious users while giving responses, and the dataset has also been manually checked in case of personal or company data. ## Considerations for Using the Data ### Social Impact of Dataset We believe that the dataset could be widely useful in the emerging field of Finnish-speaking chat applications. One of the central aims of the Tuumailubotti project is to improve the representation of neurodiversity in Large Language Models (LLMs). While we did not purposefully seek to create a dataset that would only feature under-represented rhetoricity, we tried to create conditions that were favorable to its emergence. We will now investigate whether we succeeded in our goal to create a chatbot that could allow for a wider range of communication, and whether this translates to increased accessibility of the bot itself. We also acknowledge that the data could be used for harmful scenarios, such as the identification of neurodivergent individuals. This is why we have kept information about the contributors vague on purpose. ### Discussion of Biases We tried to make the data more broadly applicable outside of YLE by purposefully avoiding company jargon. However, with 70% of the data coming from YLE employees, the data is bound to be skewed toward the kind of wofk life scenarios faced by employees of a broadcast company. Of course, the types of biases that prevent members of neurominorities entering work life also affected the types of contributors we had available. ## Additional Information ### Dataset Curators Vertti Luostarinen, technical lead of the Tuumailubotti project, AI Engineer at CasvuGen Oy and MA student in New Media in Aalto University School of Arts, Design and Architecture. His thesis work is supported by the Media Industry Research Foundation of Finland. Anni Klutas and Johan Sundström from Yle's HR Department. Marika Björn from Yle's Creative Content and Media Department. ### Licensing Information We release this dataset under the "Tuumailubotti licence". Please see the licence section for more information. ### Citation Information Vertti Luostarinen is working on his MA Thesis about the Tuumailubotti project, under the working title "Neurodiverse Narratives with Conversational AI Systems." Citation information is still pending.