cevahir-ai

Cevahir Zihin: a mind that learns by living, not a Transformer. It keeps knowledge as evidence on ties, attends with a Seven-Place Focus and learns from every c

What is cevahir-ai?

Cevahir Zihin: a mind that learns by living, not a Transformer. It keeps knowledge as evidence on ties, attends with a Seven-Place Focus and learns from every conversation. Also: a Turkish BPE tokenizer, a Transformer language-model core and cognitive systems.

Overview of cevahir-ai

What it is

Cevahir AI is an open-source research engine that includes a new mind core called Cevahir Zihin, which learns by living rather than by training. It is not a neural network; it has no weight matrices, embedding vectors, backpropagation, or softmax attention. Instead, it keeps every relation as a tie and stores evidence by source (seen, told, imitated, own act, inferred, or imagined). The repository also contains a Turkish BPE tokenizer, a Transformer language-model core, and cognitive systems.

Key features

  • Seven-Place Focus: at each moment only seven units enter the focus; there is no context window, and time per moment grows slowly (about 2 ms on a small mind, about 4 ms at 3 million ties).
  • Exact Retraction: the evidence of any single experience can be taken back exactly (in the program, !unut covers what you typed in the current session).
  • Recall from one reading: given the first two words of a random sentence, the mind said a sentence it had read that begins with them in 145 of 200 tries, and the rest of that very sentence in 60 of 61 cases where only one sentence starts with those two words.
  • Working memory (Wholes): on SCAN's add-jump split with commands of every length (7,706 test commands), it answers 7,706 of 7,706 after watching 8% of the training dialogues twice.
  • Told Apart: a word never said alone answers with what two dialogues that differ only in it differ in; SCAN's around right template split goes from 0 to 500 of 500.
  • Not knowing: it feels when it does not know (a measured inner state) and learns how to say so from the people it lives with.
  • Dreams and learning from consequences: in sleep it dreams about the day's topics; in research tests it learned from test results alone to choose the correct completion.

Use cases

Research into alternative mind architectures that learn continuously from conversation. Educational exploration of non-Transformer cognitive systems. Building a conversational agent that learns language and knowledge from interaction without a separate training phase.

Getting started

git clone https://github.com/myylogic/cevahir-aicd cevahir-aipython3 -m venv .venv && . .venv/bin/activatepip install torch numpypython -m zihinOn Windows PowerShell:git clone https://github.com/myylogic/cevahir-aicd cevahir-aipip install torch numpy. .cevahir.ps1Start inside the cevahir-ai folder. The program speaks Turkish; on first start the mind is empty and learns by living with you.

License and pricing

Licensed under Apache-2.0. Open source.

這是什麼

Cevahir AI 是一個開源研究引擎,包含一個新的核心「Cevahir Zihin」,這是一個透過生活而非訓練來學習的心智。它不是神經網路;沒有權重矩陣、嵌入向量、反向傳播或 softmax 注意力。它將每個關係儲存為連結,並根據來源(親眼所見、被告知、模仿、自身行為、推斷或想像)保留證據。此儲存庫也包含土耳其語 BPE 分詞器、Transformer 語言模型核心和認知系統。

主要功能

  • 七位焦點:每個時刻只有七個單位進入焦點;沒有上下文視窗,每個時刻的時間增長緩慢(小型心智約 2 毫秒,300 萬個連結時約 4 毫秒)。
  • 精確撤回:任何單一經驗的證據都可以精確地收回(在程式中,!unut 涵蓋你在當前會話中輸入的內容)。
  • 一次閱讀回憶:給定一個隨機句子的前兩個詞,心智在 200 次嘗試中有 145 次說出它讀過以這兩個詞開頭的句子,在 61 個只有一個句子以這兩個詞開頭的情況中,有 60 次說出該句子的其餘部分。
  • 工作記憶(整體):在 SCAN 的 add-jump 分割中,針對所有長度的指令(7,706 個測試指令),在觀看 8% 的訓練對話兩次後,它回答了 7,706 個中的 7,706 個。
  • 區分:一個從未被單獨說過的詞,會根據僅在該詞上不同的兩個對話的差異來回答;SCAN 的 around right 模板分割從 0 提升到 500 個中的 500 個。
  • 不知道:它會感覺到何時不知道(一個可測量的內部狀態),並從與之生活的人那裡學習如何表達。
  • 夢境與從後果學習:在睡眠中它會夢見當天的主題;在研究測試中,它僅從測試結果中學習選擇正確的完成方式。

使用情境

研究從對話中持續學習的替代心智架構。教育性地探索非 Transformer 的認知系統。建立一個無需單獨訓練階段即可從互動中學習語言和知識的對話代理。

快速開始

git clone https://github.com/myylogic/cevahir-aicd cevahir-aipython3 -m venv .venv && . .venv/bin/activatepip install torch numpypython -m zihin在 Windows PowerShell 上:git clone https://github.com/myylogic/cevahir-aicd cevahir-aipip install torch numpy. .cevahir.ps1在 cevahir-ai 資料夾內啟動。該程式使用土耳其語;首次啟動時心智是空的,並透過與你一起生活來學習。

授權與價格

採用 Apache-2.0 授權。開源。

Based on the project's public README and repository metadata.

Add this badge to your README

Featured on YunoDB Badge styles & help

Maintain this project? Show that it is featured on YunoDB.

Markdown[![Featured on YunoDB](https://yunodb.com/badge/myylogiccevahir-ai.svg)](https://yunodb.com/tools/myylogiccevahir-ai?utm_source=github&utm_medium=badge)
HTML<a href="https://yunodb.com/tools/myylogiccevahir-ai?utm_source=github&utm_medium=badge"><img src="https://yunodb.com/badge/myylogiccevahir-ai.svg" alt="Featured on YunoDB" height="20"></a>