| license: mit | |
| task_categories: | |
| - text-generation | |
| language: | |
| - en | |
| tags: | |
| - synthetic | |
| - instruction-tuning | |
| - alpaca-format | |
| - lora | |
| - notebookops | |
| pretty_name: NotebookOps Synthetic Instruction Dataset | |
| size_categories: | |
| - n<1K | |
| # NotebookOps Synthetic Instruction Dataset | |
| NotebookOps is a small synthetic instruction-tuning dataset for beginner | |
| fine-tuning experiments. It teaches a model to convert natural-language | |
| productivity requests into a strict fake command format. | |
| The format is intentionally artificial so that the before/after fine-tuning | |
| effect is easy to see: | |
| ```text | |
| <NBOOK> | |
| intent: reminder.create | |
| recipient: Sofia | |
| date: tomorrow | |
| time: 09:00 | |
| task: review invoice follow-up | |
| </NBOOK> | |
| ``` | |
| ## Files | |
| - `train.jsonl`: 500 generated training examples | |
| - `eval.jsonl`: 80 held-out generated evaluation examples | |
| Each row contains: | |
| - `instruction`: the instruction prompt | |
| - `input`: the natural-language request | |
| - `output`: the target NotebookOps record | |
| ## Intents | |
| The dataset covers five synthetic intents: | |
| - `reminder.create` | |
| - `email.send` | |
| - `task.create` | |
| - `note.create` | |
| - `calendar.create` | |
| ## Intended Use | |
| This dataset is meant for small LoRA/SFT demonstrations where the goal is to | |
| make fine-tuning behavior visually obvious. It is not intended as a production | |
| parser or a benchmark. | |
| ## Generation | |
| The dataset was generated deterministically with: | |
| ```bash | |
| python3 generate_notebookops_dataset.py --train-size 500 --eval-size 80 --seed 3407 | |
| ``` | |