zgsndzq/superai

How zgsndzq/superai works

SuperAI is a Python idiom-chain model and chat application: users can train and evaluate a ternary idiom engine, play chain-prediction workflows, or chat through a router that selects idiom, calculation, or general-language experts. The graph includes the documented CLI capabilities and separates router training from interactive inference. Relationships are limited to calls shown in excerpts, source-index references, or README statements; unsampled command internals are not inferred.

Main entry points: app.py Command dispatch · moe.py MoE chat and dispatch · fused_train_cmd.py Router training command · eval_cmd.py Evaluation command

Components 21

CLI and interaction

Python entry llm.py
Command dispatch cli/app.py
Base chat command cli/chat_cmd.py
Fused chat command cli/fused_chat_cmd.py
Train command cli/train_cmd.py
Test command cli/test_cmd.py
Evaluation command cli/eval_cmd.py
Router training command cli/fused_train_cmd.py

Idiom engine

Idiom vocabulary and pairs core/data.py
Idiom chain model core/model.py
Ternary linear layer core/ternary.py
Chain training and export core/trainer.py
Chain inference core/infer.py

Router and experts

MoE chat and dispatch core/moe.py
Ternary router core/ternary.py
Idiom-chain expert core/moe.py
Calculation expert core/calc.py
General Qwen expert core/moe.py

Training and evaluation

Model evaluation cli/eval_cmd.py

Other components

User
Qwen base model

Connections 29

User → Python entry: runs commands README.md
Python entry → Command dispatch: invokes llm.py
Command dispatch → Base chat command: dispatches cli/app.py
Command dispatch → Fused chat command: dispatches cli/app.py
Command dispatch → Train command: dispatches cli/app.py
Command dispatch → Test command: dispatches cli/app.py
Command dispatch → Evaluation command: dispatches cli/app.py
Command dispatch → Router training command: dispatches cli/app.py
Train command → Chain training and export: starts training cli/train_cmd.py
Chain training and export → Idiom vocabulary and pairs: loads and builds pairs core/trainer.py
Chain training and export → Idiom chain model: trains core/trainer.py
Idiom chain model → Ternary linear layer: uses layers core/model.py
Test command → Chain inference: supports play workflow README.md
Chain inference → Idiom chain model: loads checkpoint model core/infer.py
Evaluation command → Idiom vocabulary and pairs: builds evaluation pairs cli/eval_cmd.py
Evaluation command → Chain inference: loads and scores models cli/eval_cmd.py
Evaluation command → Chain training and export: uses compression stats cli/eval_cmd.py
Router training command → Idiom vocabulary and pairs: builds router samples cli/fused_train_cmd.py
Router training command → MoE chat and dispatch: constructs MoE cli/fused_train_cmd.py
Router training command → Qwen base model: encodes samples cli/fused_train_cmd.py
Router training command → Ternary router: trains router weights cli/fused_train_cmd.py
Fused chat command → MoE chat and dispatch: loads and chats cli/fused_chat_cmd.py
MoE chat and dispatch → Qwen base model: encodes and generates core/moe.py
MoE chat and dispatch → Ternary router: routes hidden state core/moe.py
MoE chat and dispatch → Idiom-chain expert: dispatches chain requests core/moe.py
MoE chat and dispatch → Calculation expert: calls calculation helpers core/moe.py
MoE chat and dispatch → General Qwen expert: falls back to generation core/moe.py
Idiom-chain expert → Chain inference: predicts valid continuations core/moe.py
Ternary router → Ternary linear layer: uses ternary layer core/moe.py

More Python diagrams: scutlihaoyu/open-chat-video-editor · sigma67/ytmusicapi · awslabs/amazon-redshift-utils · tencent-hunyuan/hunyuanworld-1.0 · google-research/tabfm

Drawn by GitDiagram on September 30, 2026 from the repository's file tree, README and source files. Read as Markdown ·

How zgsndzq/superai works

Main entry points: app.py Command dispatch · moe.py MoE chat and dispatch · fused_train_cmd.py Router training command · eval_cmd.py Evaluation command

Connections 29

User → Python entry: runs commands README.md
Python entry → Command dispatch: invokes llm.py
Command dispatch → Base chat command: dispatches cli/app.py
Command dispatch → Fused chat command: dispatches cli/app.py
Command dispatch → Train command: dispatches cli/app.py
Command dispatch → Test command: dispatches cli/app.py
Command dispatch → Evaluation command: dispatches cli/app.py
Command dispatch → Router training command: dispatches cli/app.py
Train command → Chain training and export: starts training cli/train_cmd.py
Chain training and export → Idiom vocabulary and pairs: loads and builds pairs core/trainer.py
Chain training and export → Idiom chain model: trains core/trainer.py
Idiom chain model → Ternary linear layer: uses layers core/model.py
Test command → Chain inference: supports play workflow README.md
Chain inference → Idiom chain model: loads checkpoint model core/infer.py
Evaluation command → Idiom vocabulary and pairs: builds evaluation pairs cli/eval_cmd.py
Evaluation command → Chain inference: loads and scores models cli/eval_cmd.py
Evaluation command → Chain training and export: uses compression stats cli/eval_cmd.py
Router training command → Idiom vocabulary and pairs: builds router samples cli/fused_train_cmd.py
Router training command → MoE chat and dispatch: constructs MoE cli/fused_train_cmd.py
Router training command → Qwen base model: encodes samples cli/fused_train_cmd.py
Router training command → Ternary router: trains router weights cli/fused_train_cmd.py
Fused chat command → MoE chat and dispatch: loads and chats cli/fused_chat_cmd.py
MoE chat and dispatch → Qwen base model: encodes and generates core/moe.py
MoE chat and dispatch → Ternary router: routes hidden state core/moe.py
MoE chat and dispatch → Idiom-chain expert: dispatches chain requests core/moe.py
MoE chat and dispatch → Calculation expert: calls calculation helpers core/moe.py
MoE chat and dispatch → General Qwen expert: falls back to generation core/moe.py
Idiom-chain expert → Chain inference: predicts valid continuations core/moe.py
Ternary router → Ternary linear layer: uses ternary layer core/moe.py

Drawn by GitDiagram on September 30, 2026 from the repository's file tree, README and source files. Read as Markdown ·