Snapkitty/summon

╔══════════════════════════════════════════════════════════════════════════════╗
║                                                                              ║
║   ⬡  S U M M O N                                                            ║
║                                                                              ║
║   ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓                                ║
║   Build your own weights. Name your own model.                               ║
║   Sovereign fine-tuning framework · pip install summon                       ║
║   ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓                                ║
║                                                                              ║
║   ⬡ Ω ↺ Ψ Δ Λ Σ Φ α  — WORM SEALED AT EVERY STEP                          ║
║                                                                              ║
╚══════════════════════════════════════════════════════════════════════════════╝

Summon is a pip-installable Python framework for building sovereign fine-tuned language models. It wraps HuggingFace PEFT + TRL + bitsandbytes into a single fluent chain: Summon.begin("YourModel") → .base() → .corpus() → .constitutional() → .license() → .train() → .push(). The API is designed in the functional / immutable style — every method returns a new SovereignModel instance; nothing mutates in place. The Corpus builder is equally composable: stack named layers from JSONL files or raw string lists, then .seal() to freeze them. A SHA-256 WORM chain runs through every step of both pipelines, producing a verifiable manifest of exactly what data, what base, and what constitution went into your weights. Training is QLoRA (4-bit NF4 quantization via bitsandbytes, r=16 lora_alpha=32, target modules q_proj/v_proj/k_proj/o_proj) using HuggingFace SFTTrainer. Supported base models include Nemotron Mini 4B, Llama 3 8B/70B, Mistral 7B, Phi-3 Mini, Qwen2 7B, Gemma2 9B, and Falcon 7B — or pass any HuggingFace ID directly.

Architecture

flowchart LR
    subgraph Corpus Builder
        C0([Corpus.layer 0\ngenesis.jsonl])
        C1([Corpus.layer 1\nenochian.jsonl])
        C2([Corpus.layer N\n...])
        CS([.seal\nWORM hash])
        C0 --> C1 --> C2 --> CS
    end

    subgraph SovereignModel Chain
        M0([Summon.begin\nname]) --> M1
        M1([.base\nnominate HF model]) --> M2
        M2([.corpus\nlayers / Corpus obj]) --> M3
        M3([.constitutional\nprinciples list]) --> M4
        M4([.license\nsovereign-source-v1]) --> M5
        M5([.train\nQLoRA 4-bit]) --> M6
        M6([.push\nHuggingFace Hub])
    end

    CS -->|corpus_obj| M2

    subgraph Trainer
        T0[Load base model\nBitsAndBytesConfig NF4]
        T1[Apply LoraConfig\nr=16 alpha=32]
        T2[SFTTrainer\nepochs · batch · lr]
        T3[save_model\nwrite model card]
        T0 --> T1 --> T2 --> T3
    end

    M5 --> Trainer

    subgraph WORM Chain
        W0[GENESIS] --> W1[BASE seal]
        W1 --> W2[CORPUS seal]
        W2 --> W3[CONSTITUTION seal]
        W3 --> W4[LICENSE seal]
        W4 --> W5[TRAIN seal]
        W5 --> W6[PUSH seal]
    end

    M1 -.->|SHA-256| W1
    M2 -.->|SHA-256| W2
    M3 -.->|SHA-256| W3
    M4 -.->|SHA-256| W4
    M5 -.->|SHA-256| W5
    M6 -.->|SHA-256| W6

File Tree

summon/
├── summon/
│   ├── __init__.py             # Summon class — .begin() and .corpus() entry points
│   ├── identity/
│   │   ├── __init__.py
│   │   └── model.py            # SovereignModel — fluent chain (.base/.corpus/.constitutional/.license/.train/.push/.manifest)
│   ├── corpus/
│   │   ├── __init__.py
│   │   └── builder.py          # Corpus — layered JSONL builder (.layer/.layer_raw/.seal/.export/.summary)
│   └── train/
│       ├── __init__.py
│       └── runner.py           # Trainer — QLoRA fine-tuning (validate/run/_write_model_card)
├── examples/
│   └── build_my_model.py       # Three usage patterns: full pipeline, corpus-first, dry run
├── setup.py                    # pip packaging (extras: [train] and [hub])
└── README.md

Quick Start

# Install (core — no GPU deps)
pip install summon

# Install with training deps
pip install "summon[train]"          # torch, transformers, peft, trl, datasets, bitsandbytes, accelerate

# Install with HuggingFace Hub push support
pip install "summon[train,hub]"

Full pipeline:

from summon import Summon

model = (
    Summon.begin("AhmadMeta-v1")
    .base("nemotron-mini-4b")           # or full HF ID: "nvidia/Minitron-4B-Base"
    .corpus(layers=[
        "data/the_book.jsonl",
        "data/enoch.jsonl",
        "data/circle7.jsonl",
    ])
    .constitutional(["truth", "sovereignty", "evidence", "no_deception"])
    .license("sovereign-source-v1")
    .train(device="cuda", epochs=3, batch_size=4, learning_rate=2e-4)
    .push("my-org/AhmadMeta-v1")
)

model.manifest()    # print WORM-sealed manifest, optionally write to file

Corpus builder separately:

from summon import Summon

corpus = (
    Summon.corpus()
    .layer(0, "data/genesis.jsonl",   name="genesis")
    .layer(1, "data/enoch.jsonl",     name="enochian")
    .layer_raw(2, ["raw text line 1", "raw text line 2"], name="inline")
    .seal()
)

model = (
    Summon.begin("JessicaLM-v1")
    .base("llama3-8b")
    .corpus(corpus_obj=corpus)       # pass the sealed Corpus object
    .constitutional(["truth", "care", "sovereignty"])
    .license("apache-2.0")
    .train(device="cuda", epochs=5)
    .push("jessica-org/JessicaLM-v1")
)

Dry run (validate config, no GPU):

model = (
    Summon.begin("TestModel-v1")
    .base("phi3-mini")
    .corpus(layers=["data/sample.jsonl"])
    .constitutional(["truth"])
    .train(dry_run=True)    # validates, does not launch training
)

Key Features

• Single fluent chain — Summon.begin("name").base().corpus().constitutional().license().train().push() — the entire pipeline in one expression
• Functional / immutable API — every method returns a new SovereignModel instance; state never mutates; safe to branch at any point
• Layered Corpus builder — stack named JSONL layers with .layer() (from file) or .layer_raw() (from string list); .seal() freezes and WORM-stamps the corpus; .export() merges all layers to a single JSONL for training
• QLoRA fine-tuning — 4-bit NF4 quantization via bitsandbytes, LoRA rank 16, alpha 32, targets q_proj/v_proj/k_proj/o_proj, SFTTrainer with gradient accumulation and fp16; auto-writes a HuggingFace model card with WORM seal
• Constitutional principles — .constitutional(["truth","sovereignty","evidence"]) bakes your principles into the model manifest and model card; shapes RLHF/preference data generation
• Eight supported base models — Nemotron Mini 4B, Llama 3 8B/70B, Mistral 7B, Phi-3 Mini, Qwen2 7B, Gemma2 9B, Falcon 7B — or pass any HuggingFace model ID directly
• WORM chain on every step — SHA-256 hash chain from GENESIS through BASE → CORPUS → CONSTITUTION → LICENSE → TRAIN → PUSH; final hash in .manifest() proves exact provenance
• .manifest() output — prints and optionally writes a JSON document with name, base, corpus layers, constitution, license, output path, WORM head hash, and creation timestamp
• HuggingFace Hub push — .push("org/model") calls HfApi().upload_folder() with optional private=True; warns if .train() was not called first

---

Apache 2.0 · SnapKitty Collective 2026 · Evidence or Silence