| --- |
| dataset_info: |
| features: |
| - name: conversation |
| list: |
| - name: role |
| dtype: string |
| - name: text |
| dtype: string |
| splits: |
| - name: train |
| num_bytes: 8022604.792326268 |
| num_examples: 3721 |
| download_size: 4325950 |
| dataset_size: 8022604.792326268 |
| license: apache-2.0 |
| language: |
| - de |
| size_categories: |
| - 1K<n<10K |
| --- |
| # German OpenAssistant Conversations Dataset (OASST-DE) |
| With the goal of advancing open-source, german-language LLM research, we present |
| OASST-DE: a high quality subset of a recent (25.08.23) dump from the [OpenAssistant website](https://www.open-assistant.io/) |
| translated to German using the GPT-3.5 API. More details on how the dataset was filtered and translated under [dataset creation.](#dataset-creation-process) |
| For more details on the OpenAssistant Project, look at the [first OASST dataset (OASST1)](https://huggingface.co/datasets/OpenAssistant/oasst1), [the Open-Assistant GitHub repo](https://github.com/LAION-AI/Open-Assistant) |
| or [our paper](https://arxiv.org/abs/2304.07327). |
|
|
| This dataset was created as part of LAION's LeoLM (Linguistically Enhanced Open Language Model) project led by Björn Plüster. |
| Check out LeoLM-Chat trained with OASST-DE ([7b](https://huggingface.co/LeoLM/leo-hessianai-7b-chat), [13b](https://huggingface.co/LeoLM/leo-hessianai-13b-chat)) finetuned on OASST-DE and read [their blog post](https://laion.ai/blog/leo-lm/)) for more info on LeoLM. |
|
|
| ## Dataset Creation Process |
| This dataset was created from a recent OASST dump by following these steps: |
| - Filter for Top1 response trees with assistant response leaves |
| - Filter first prompt quality >= 0.5 |
| - Filter total conversation length < 1900 tokens to fit in GPT3.5 context length |
| - Filter for `'lang' == 'de'` -> add to dataset |
| - Filter for `'lang' == 'en'` (other languages often result in failed translations) |
| - Translate using GPT-3.5-turbo API (total cost ~15$). |
|
|
| This results in around 3.7k samples of high-quality assistant conversations. |
|
|
| ## Dataset Structure |
| This dataset has only one `'conversation'` field. Each example is a list of an alternating conversation between `'prompter'` and `'assistant'`, |
| where each entry is a dict with `'text'` and `'role'` fields: |
| ```json |
| "conversation": [ |
| {"role": "prompter", "text": "Moin, wie geht's dir?"}, |
| {"role": "assistant", "text": "Moin Moin! Mir geht es gut, und dir?"}, |
| ... |
| ] |
| ``` |
|
|
| ## Usage with 🤗Datasets: |
| ```python |
| from datasets import load_dataset |
| ds = load_dataset("OpenAssistant/OASST-DE", split="train") |
| print(ds[0]["conversation"]) |
| ``` |
