---
language:
- en
license: apache-2.0
task_categories:
- token-classification
tags:
- prompt-compression
- disfluency
- synthetic-data
size_categories:
- 100K<n<1M
pretty_name: Saccade 100K
---

# Saccade 100K

A 100,000-example word-level `KEEP`/`DROP` dataset for experimental prompt
compression.

## Composition

| Split | Examples |
|---|---:|
| Train | 80,000 |
| Validation | 10,000 |
| Test | 10,000 |

The source texts are English, human-written `prompter` messages selected from
the Apache-2.0 licensed OpenAssistant OASST1 dataset.

The labels are generated by synthetically inserting:

- hesitation and filler phrases,
- redundant politeness,
- repeated spans,
- false starts and verbal restarts.

Original source words receive `KEEP`; inserted words receive `DROP`.

## Important limitation

The source prompts are human-written, but the disfluencies and labels are
synthetically generated. This is not a manually annotated corpus and should not
be represented as one.

## Fields

- `tokens`: word-level noisy input
- `labels`: `0 = DROP`, `1 = KEEP`
- `source_text`: original human-written prompt
- `noisy_text`: reconstructed noisy prompt