| language: | |
| - en | |
| license: apache-2.0 | |
| library_name: transformers | |
| pipeline_tag: token-classification | |
| tags: | |
| - prompt-compression | |
| - token-classification | |
| - tiny-model | |
| base_model: google/bert_uncased_L-2_H-128_A-2 | |
| # Saccade Tiny | |
| Saccade Tiny is a small experimental token classifier that predicts whether | |
| each word in a prompt should be **KEEP** or **DROP** before LLM inference. | |
| ## Architecture | |
| - Base model: `google/bert_uncased_L-2_H-128_A-2` | |
| - Task: binary token classification | |
| - Labels: `KEEP`, `DROP` | |
| ## Training data | |
| This first version was trained primarily on synthetic examples. Clean | |
| instructions were modified with injected filler and redundant wording. | |
| Original words were labelled `KEEP`; injected words were labelled `DROP`. | |
| ## Held-out evaluation | |
| - Accuracy: 1.0000 | |
| - KEEP precision: 1.0000 | |
| - KEEP recall: 1.0000 | |
| - KEEP F1: 1.0000 | |
| ## Intended use | |
| Research and demonstrations involving conservative filler removal from English | |
| prompts. | |
| ## Limitations | |
| The model was trained mainly on synthetic data and may remove meaningful | |
| context. It has not yet demonstrated a general improvement in downstream LLM | |
| accuracy. Do not use it for safety-critical, legal, medical, or financial text. | |