Hui

Hui means gathering/assembly. Hui lets you convene any models, debate clearly, and grow new skills fast.

Architecture (signature)

flowchart TB
  subgraph "Monorepo: hui"
    direction TB
    subgraph apps["apps/"]
      direction LR
      subgraph collab["apps/collab (Hui Council: real-time model council)"]
        direction TB
        UI["Gradio UI\\nSwitchboard (models) + Role Foundry (roles)"]
        VCE["Consensus Engine"]
        UI --> VCE
      end
      subgraph finetune["apps/finetune (Hui Train: repurpose & fine-tune)"]
        direction TB
        CLUI["CLI/Notebook Runner\\n(ingest → prepare → train → eval → package)"]
        Pipelines["Pipelines: JSONL prep • SFT/LoRA/Unsloth • Eval • Packaging"]
        CLUI --> Pipelines
      end
    end

    subgraph packages["packages/"]
      direction LR
      subgraph core["packages/core"]
        direction TB
        RA["EnhancedResearchAgent\\n(cache • coalesce • parallel)"]
        Tools["Tools: web • wikipedia • arxiv • github • sec"]
        Cache["Caches: LRU + in-flight coalescing + KV hooks"]
        DSPy["DSPy Synthesis (optional)"]
        Reports["Structured Report Templates"]
        Logger["Dataset Logger (JSONL)"]
        RA --- Tools
        RA --- Cache
        RA --- Reports
        RA --- Logger
        VCE === RA
      end
      subgraph adapters["packages/adapters"]
        direction TB
        Plug["Model Plugboard\\n(OpenAI-compatible endpoints)"]
        HuiIO["Hui IO Adapter\\n(ingest tasks, transcripts, outputs)"]
        VCE --- Plug
        Pipelines --- HuiIO
      end
      subgraph training["packages/training"]
        direction TB
        Prep["Data Prep: normalize • schema • dedupe • splits"]
        Train["Trainer: SFT/LoRA • Unsloth compatible"]
        Eval["Eval: regression sets • rubric • structured report checks"]
        Pack["Packaging: checkpoints • prompt templates • adapters"]
        Pipelines === Prep
        Prep --> Train --> Eval --> Pack
      end
    end

    subgraph data["data/"]
      direction TB
      JSONL["JSONL Traces (SFT-ready)"]
      Artifacts["Artifacts (models/checkpoints)"]
      ReportsOut["Generated Reports (structured)"]
      Logger --- JSONL
      Pack --- Artifacts
      CLUI --- ReportsOut
    end
  end

  subgraph ext["External Services"]
    direction LR
    Mistral["Mistral API"]
    Samba["SambaNova API"]
    OpenAICompat["Other OpenAI-Compat Endpoints"]
  end

  Plug --- OpenAICompat
  VCE --- Mistral
  VCE --- Samba

  subgraph wf1["Workflow A: Hui Council"]
    direction TB
    A1["Switchboard: add models"]
    A2["Role Foundry: assign roles"]
    A3["Debate: N rounds + live research"]
    A4["Final synthesis (structured)"]
    A5["Optional: log traces → JSONL"]
    UI --> A1 --> A2 --> A3 --> A4 --> A5 --> JSONL
  end

  subgraph wf2["Workflow B: Hui Train"]
    direction TB
    B1["Ingest via Hui IO"]
    B2["JSONL prep"]
    B3["Train (Unsloth)"]
    B4["Evaluate (rubric)"]
    B5["Package → Plugboard"]
    HuiIO --> B1 --> B2 --> JSONL
    JSONL --> B3 --> B4 --> B5 --> Artifacts
  end

Loading

What Hui enables

• Real‑time ANY‑model collaboration via Switchboard and Role Foundry
• Deterministic structured reports for high‑stakes outputs
• Rapid fine‑tuning/repurposing using SFT/LoRA (Unsloth‑friendly)
• Reusable data pipeline: ingest → JSONL → train → evaluate → package

Monorepo layout

hui/
  apps/
    collab/           # Hui Council UI (Switchboard, Role Foundry)
    finetune/         # Hui Train CLI/Notebooks
  packages/
    core/             # Research agent, templates, logging, DSPy (opt)
    adapters/         # Plugboard + IO adapters
    training/         # Prep/Train/Eval/Package utilities
  data/               # JSONL traces, artifacts, reports

Quickstart

• Council: python apps/collab/app.py
• Train:
  • python apps/finetune/cli.py ingest input_dir_or_jsonl data/raw.jsonl
  • python apps/finetune/cli.py prepare data/raw.jsonl data/prepared.jsonl
  • python apps/finetune/cli.py evaluate data/prepared.jsonl reports/eval.json --type enhancement_plan_v1 --threshold 0.7
  • python apps/finetune/cli.py unsloth-cmd data/prepared.jsonl models/hui-lora --base meta-llama/Meta-Llama-3.1-8B-Instruct