vynly/CAEL-0.3.4

CAEL 0.3.4

CAEL is a hybrid non-neural assistant: counted language models, cellular automata, symbolic reasoning, exact arithmetic, a bounded Python interpreter, local retrieval and hand-written knowledge. No neural network, transformer or external language model runs at answer time. This is a custom Python runtime, not a Transformers/AutoModel checkpoint. Source: GitHub, Vovala14/CAEL, or CAEL-0.3.4.zip here. For an offline install, put cael-0.3.4.cael in the source folder's models/ directory and run python -m cael_unified --download.

CAEL 0.3.4 understands misspellings, answers far more everyday questions, writes everyday Python, reasons through everyday problems, and plays Sudoku. Every runtime engine is still non-neural.

Install

git clone https://github.com/Vovala14/CAEL.git
cd CAEL
python -m pip install -r requirements-unified.txt
python -m cael_unified --download   # fetches, verifies and installs cael-0.3.4.cael (~561 MB)
python -m cael_unified --ui         # http://127.0.0.1:8765

What's new

• Spelling and chat shorthand. "whats the captial of australia", "wich planet", "r u smart" and "ur funny" are read as meant.
  • It uses word and word-pair counts from CAEL's Wikipedia snapshot and chat corpus (new models/cael-spelling-v1, CC BY-SA 4.0), plus WordNet.
  • The neighbouring words decide between candidate corrections.
  • Names, code and correct text are left alone.
• Fewer refusals. New answers for:
  • abbreviations and job descriptions;
  • stories about any subject;
  • birthday, thank-you and get-well messages, and pet names;
  • tips, riddles and compliments;
  • more everyday facts, a country's languages, and "where is X" with the city.
• Everyday Python.
  • 60 tested functions for common requests.
  • A composer that builds small functions from the request's wording.
  • A guide to Python basics.
  • Every function is run on its examples before it is shown.
• Everyday reasoning. Each answer shows its working.
  • Word problems and "at the same time" trick questions.
  • Days, clock times, ages, comparing numbers and counting letters.
  • Sequences and odd one out.
  • Where things are, family relations and "all/some" logic.
• Sudoku. "play sudoku" (easy, medium, hard, expert or highest) gives a new puzzle with exactly one solution.
  • Play it move by move in the chat: r3c4=7, "put 7 in row 3 column 4", undo, check, hint, solve it.
  • It solves any pasted grid, including the hardest known puzzles, in milliseconds (Knuth's Algorithm X).
  • The highest level is a reshuffled famous hardest puzzle with the same difficulty.

Results on fresh questions written after the changes

	Before	0.3.4
Everyday questions, spelled correctly	85%	93%
Everyday questions, misspelled	35%	87%
Python questions	5/10	10/10
Reasoning questions	3/20	19/20
Function requests (last untouched set)	–	10/10

Files

• cael-0.3.4.cael: the text-only model bundle (now with the spelling counts).
• CAEL-0.3.4.zip: the source, exactly as committed.

Licence: CAEL Non-Commercial License — free for personal and research use; commercial use needs a separate licence. Third-party parts keep their own licences (NOTICE.md); the spelling counts are CC BY-SA 4.0.

Results are local, maintainer-run measurements on question sets written for this release; they are not a benchmark leaderboard.