README.md

---
language:
- en
license: apache-2.0
library_name: transformers
pipeline_tag: token-classification
tags:
- prompt-compression
- token-classification
- tiny-model
- synthetic-data
base_model: google/bert_uncased_L-2_H-128_A-2
datasets:
- aayushbist/saccade-100k
---

# Saccade Tiny 100K

Saccade Tiny is a **4,369,666-parameter** experimental token
classifier that predicts whether each word should be `KEEP` or `DROP` before a
prompt is sent to a larger language model.

## Training

- Base model: `google/bert_uncased_L-2_H-128_A-2`
- Training examples: **80,000**
- Validation examples: **10,000**
- Held-out test examples: **10,000**
- Total generated dataset: **100,000 examples**
- Dataset: [aayushbist/saccade-100k](https://huggingface.co/datasets/aayushbist/saccade-100k)

The examples were derived from human-written English OpenAssistant prompts by
injecting synthetic filler, repetitions, false starts, and redundant
politeness.

## Held-out synthetic evaluation

| Metric | Result |
|---|---:|
| Accuracy | 0.9393 |
| KEEP precision | 0.9363 |
| KEEP recall | 0.9684 |
| KEEP F1 | 0.9521 |
| Predicted DROP rate | 35.66% |

## Intended use

Research and demonstrations involving conservative removal of obvious
low-information wording from English prompts.

## Honest limitations

- The source prompts are human-written, but the noise and labels are synthetic.
- High performance on this test set does not prove equivalent performance on
  naturally disfluent speech or unseen domains.
- This version has not established a general increase in downstream LLM
  accuracy.
- Do not use it for safety-critical, legal, medical, or financial text without
  extensive additional evaluation.

## Usage

```python
from transformers import pipeline

classifier = pipeline(
    "token-classification",
    model="aayushbist/saccade-tiny-100k",
    aggregation_strategy="none",
)

print(classifier("Could you basically summarize this report?"))
```