| | --- |
| | dataset_info: |
| | features: |
| | - name: text |
| | dtype: string |
| | splits: |
| | - name: train |
| | num_bytes: 800969 |
| | num_examples: 647 |
| | download_size: 456786 |
| | dataset_size: 800969 |
| | configs: |
| | - config_name: default |
| | data_files: |
| | - split: train |
| | path: data/train-* |
| | license: apache-2.0 |
| | language: |
| | - en |
| | size_categories: |
| | - n<1K |
| | task_categories: |
| | - text-generation |
| | --- |
| | |
| | # High-Quality OpenAssistant Subset for Reverse Instruction Generation |
| |
|
| | ## Dataset Description |
| | This dataset is a carefully curated subset of the [Open Assistant dataset](https://huggingface.co/datasets/OpenAssistant/oasst1/tree/main), specifically designed for training models to generate instructions based on responses. |
| |
|
| | ### Key Features: |
| | - Contains only the highest-rated conversation paths from the original dataset |
| | - Filtered to include only English language conversations |
| | - Removed mentions of Open Assistant to improve generalizability |
| | - Single-turn conversations only |
| | - Formatted with Llama 2 chat model prompt instruction tokens |
| | - Response and instruction roles are swapped for reverse fine-tuning |
| |
|
| | ## Usage |
| | This dataset is designed for fine-tuning language models to generate instructions based on given responses. It has been successfully used to train Guanaco with QLoRA. |
| |
|
| | ## License |
| | This dataset is released under the Apache 2.0 license. |
