A rewrite has to earn its place.
FLAN-T5-base proposes the shorter prompt. Python checks whether to accept it.
Keep critical blocks
Mask code, JSON, URLs and explicitly protected sections.
Compress the wording
Apply conservative cleanup, then an opt-in LoRA rewriter.
Check what survived
Restore blocks; check required phrases, numbers, tokens and similarity.
Accept or fall back
Return the shorter candidate if checks pass. Otherwise keep the original.
The transformation we’re training for
An authored training pair. This is the intended behavior, not a live model prediction.
Could you please help me write a Python function that takes a list of integers as input and returns only the numbers from that list which are even numbers? Please provide the Python code.
Write a Python function that takes a list of integers and returns only the even numbers.
An honest starting point
The adapter was trained on 9 synthetic prompts, with 2 validation examples and 2 reserved test examples. The separate selector dataset contains 384 synthetic prompts.
Current validation: 0 of 2 rewrites accepted; 0% token reduction.
Both candidates lost required information, so the pipeline returned the originals. Downstream answer quality is untested. This release demonstrates the workflow and needs substantially more reviewed training data.
The goal is to reduce tokens while retaining intent, constraints and facts. Automatic checks cannot guarantee meaning preservation. This page is a static overview; it does not run inference.
Explore, reproduce, improve
Three repositories, one shared experiment.