Escarda-Rewrite (86M)

Escarda-Rewrite is an 86M-parameter prompt optimizer. Given a raw user question it rewrites it into a detailed, role-grounded instruction prompt for a downstream LLM — a persona declaration, a clear task framing, and clarifying questions.

It is part of the Escarda line (JEPA SpikeWhale backbone, vocab 16,512) and is designed to pair with Quazim0t0/Escarda-86M-Identity as the downstream answerer: user query → Escarda-Rewrite → optimized prompt → Escarda-Identity → answer.

Example

Raw query	Escarda-Rewrite output
How do I lower my cholesterol?	As a Senior Medical Expert, your task is to provide clear, concise guidance on lowering cholesterol… Begin by declaring your persona… Next, ask clarifying questions…
How do I grow tomatoes?	As an Agricultural Scientist, your task is to provide clear guidance on growing tomatoes… Begin by introducing yourself…
What are my rights if I am fired?	As a Senior Legal Advisor, your task is to provide guidance on… Begin by declaring your persona…

Usage

from rewrite import EscardaRewrite
rw = EscardaRewrite(".")            # loads model + tokenizer from this repo
print(rw.rewrite("How do I lower my cholesterol?"))

How it was built

Escarda-Rewrite was trained to match, and is benchmarked against, the prompt-optimization behaviour of QueryShield-1.5B (ml-intern-explorers/queryshield-1.5b) — a 1.5B prompt optimizer — at ~1/17th the size, using a balanced, topic-roled corpus so the chosen expert persona tracks the topic of the query (Medical, Legal, Financial, Agricultural, Software, Data Science, Marketing, Education, Research, …).

Citation

Created by Dean Byrne (Quazim0t0).

@misc{byrne2026escardarewrite,
  title  = {Escarda-Rewrite: an 86M prompt optimizer},
  author = {Byrne, Dean},
  year   = {2026}
}