CAEL 0.3.1: smarter, text-only, CAEL licence
CAEL is a hybrid non-neural assistant. It combines counted language models, cellular automata, symbolic reasoning, exact arithmetic, a bounded Python interpreter, local retrieval and hand-written knowledge. It uses no neural network, transformer or external language model when it runs. This is a custom Python runtime, not a Transformers/AutoModel checkpoint.
Download and run
- Get the source: GitHub, Vovala14/CAEL, or CAEL-0.3.1.zip from this repository.
- Install and fetch the checksum-verified 553 MB text-only bundle:
python -m pip install -r requirements-unified.txt
python -m cael_unified --download
python -m cael_unified "What is 15% of 240? Then tell me if the result is a prime number" --mode chat
python -m cael_unified --ui # http://127.0.0.1:8765
On Windows, SETUP.bat and then RUN.bat do the same. For an offline install, put
cael-0.3.1.cael in the source folder's
models/ directory and run python -m cael_unified --download.
What's new in 0.3.1
- Everyday maths with steps:
- powers, fractions as decimals or percentages;
- unit and temperature conversions;
- average speed and "3 times as many";
- shopping totals and change, and "% off".
- World facts:
- a hand-written table of well-known facts;
- a Wikipedia answer finder for superlatives, counts and years, which accepts only an opening sentence that really says it and links the article;
- capitals from Wikidata.
- Everyday writing:
- one-sentence summaries;
- emails for common purposes;
- name ideas;
- "explain X simply" and "difference between X and Y".
- Language and reasoning:
- fill-the-gap for confusable words;
- passive phrasing and better opposites;
- present-tense stories and hand-overs;
- common text and list functions in code, checked by CAEL's own interpreter.
- Multi-step requests chain all of these ("What is 2 to the power of 10? Then divide it by 4").
Results (local, maintainer-run)
| Test | 0.3 code | 0.3.1 |
|---|---|---|
| 76-prompt neutral set (Qwen2.5-0.5B-Instruct scored 65) | 56 | 75 |
| Fresh 50-prompt set written before the changes | 24 | 49 |
| Labelled chat / vocabulary / grammar / reasoning set | 102 / 102 | 102 / 102 |
| Held-out advice questions (110) | 0 wrong | 0 wrong |
How much to trust these numbers:
- The neutral set guided the work. Treat its 75 as a development score.
- The fresh set was written before any change. Its failures were not inspected while building, but its 8 facts items and 3 code items are covered by tables written afterwards. Without those 11, the score is 38 of 39.
- Earlier multiple-choice benchmark results that used the ARC Corpus are not reproduced by this release (see below).
Licence and data
- CAEL Non-Commercial License (LICENSE): free for personal and research use. Commercial use needs a separate written licence from the copyright holder (Vovala14).
- CAEL only uses third-party material that allows commercial use. Removed in 0.3.1:
- The 0.2.1 small-talk reply models, which were fitted on EmpatheticDialogues (CC BY-NC 4.0).
- The ARC Corpus evidence in the multiple-choice chooser. Its terms forbid redistribution, so it is off and its weight is 0.
- Components derived from CC BY-SA sources (Wikipedia, Stack Exchange, Dolly, Wikibooks, Wikivoyage) remain under CC BY-SA. Every model folder in the bundle lists its sources and licences, and NOTICE.md records the provenance.
- Earlier releases, including vynly/CAEL-0.3, carried a copy of the MICA licence text by mistake. CAEL is a separate model family with the same copyright holder.
Limits
- The facts table is finite. The Wikipedia finder answers only clear question shapes and says nothing rather than guess.
- Summaries are extractive and emails are templates.
- Supported reasoning and programming forms are narrow; this is not general Python understanding.
- No picture generation in 0.3.1.