This is a roadmap of basic features that are needed to make AutoCog (and STA) usable. It only considers a few weeks worth of work but I rarely have brain-cycles to work on this project.
Given that I am currently working alone on this project, I am not tracking work using issues and milestones.
In the roadmap below, each minor versions integrates the incremental progress of the previous ones.
Simply, all the v0.4.X are steps toward v0.5.
These bugfix level milestones are subject to reordering (change of priority) and shifting (introducing new milestone or actual bugfixes).
| Version | Features | Notes | Tracking |
|---|---|---|---|
| v0.4 | Structured Thoughts | release 1st version of ST | |
| v0.4.1 | Tests & Fixes | Testing more LLMs and fix tokenizations issue | |
| v0.4.2 | Roadmap & Doc | Needed some organization... | |
| v0.4.3 | FTA: Simplify, Choice Limit, and Norms | ||
| v0.4.4 | Beam Search | Implementation within FTA | |
| v0.4.5 | Python Cogs | Cog from Python files for tools | |
| v0.4.6 | PDF Reader | Read a PDF piece-by-piece and write a summary | |
| v0.5 | Language Docs | Description of the language | |
| v0.5.1 | Tests & Fixes | Expecting that it will be needed... | |
| v0.5.2 | Low-Level llama-cpp-python | ||
| v0.5.3 | Unified FTA | FTA in one "loop" using llama-cpp-python low-level API | |
| v0.5.4 | Elementary | Library of elementary "worksheet" (arithmetic: add/mul/div, literacy: spelling, grammar, comprehension) | |
| v0.5.5 | MMLU-Exams | Library of MCQ Solver using different Thought Patterns | |
| v0.5.6 | FTA to BNF | Translate FTA to llama.cpp BNF | |
| v0.6 | Tutorials | Elementary Reader & Writer | |
| v0.6.1 | HuggingFace Tramsformers | FTA implementation for HGTF | |
| v0.7 | Benchmarking | Evaluate speed and accuracy on Elementary and MMLU-Exams | |
| v0.7.1 | Tooling Benchmark | ||
| v0.8 | Finetuning | Selected foundation LLMs targetting improved performance at MMLU-Exams | |
| v0.8.1 | Finetune Tooling |