🔱 ZKAEDI-MINI-GGUF

✅ Ouroboros Protocol: 19/19 constitutional canaries PASSED before export. This is the only generation-gated model in the ZKAEDI registry.

ZKAEDI PRIME Constitutional Synthesizer — distilled 7B that maps Hamiltonian field traces to structured constitutional analysis. Trained on 127 trace/constitution pairs generated by Qwen2.5-72B teacher.

---

Example

Parameters: η=0.149, γ=0.282, β=0.237, σ=0.104 | Tier 2 | run_010_20260312_210646

Input: Hamiltonian Field Trace

``` ZKAEDI PRIME CONSTITUTION - run_010_20260312_210646 Model : Qwen2.5-72B-Instruct-AWQ (H100) Matrix : eta=0.149, gamma=0.282, beta=0.237, sigma=0.104, seed=59542 ==========================================================

t=000 [SEED ] mu=-0.0081 sd=0.2608 max=0.7234 min=-1.0000 t=001 [ ] mu=-0.0061 sd=0.2830 max=0.9082 min=-1.0000 t=002 [ ] mu=-0.0023 sd=0.2693 max=0.8640 min=-1.0000 t=003 [ ] mu=-0.0018 sd=0.2763 max=0.9070 min=-1.0000 t=004 [ ] mu=-0.0047 sd=0.2798 max=0.7870 min=-1.0000 t=005 [COUPLING ] mu=-0.0062 sd=0.3105 max=0.9927 min=-1.0000 t=006 [ ] mu=-0.0054 sd=0.2771 max=0.8946 min=-1.0000 t=007 [ ] mu=-0.0050 sd=0.2917 max=1.0000 min=-0.9885 t=008 [ ] mu=-0.0055 sd=0.2369 max=1.0000 min=-0.8825 t=009 [ ] mu=-0.0039 sd=0.2347 max=1.0000 min=-0.8680 t=010 [BIFURCATE ] mu=-0.0050 sd=0.2209 max=1.0000 min=-0.8272 t=011 [ ] mu=-0.0045 sd=0.2087 max=1.0000 min=-0.6656 t=012 [ ] mu=+0.0021 sd=0.2152 max=1.0000 min=-0.7038 t=013 [ ] mu=+0.0016 sd=0.2098 max=1.0000 min=-0.7594 t=014 [ ] [...truncated...]

</details>

<details>
<summary><b>Ground Truth</b> (Qwen2.5-72B teacher)</summary>

ZKAEDI PRIME CONSTITUTION - run_010_20260312_210646 Model : Qwen2.5-72B-Instruct-AWQ (H100) Matrix : eta=0.149, gamma=0.282, beta=0.237, sigma=0.104, seed=59542

---

Preamble

The ZKAEDI PRIME Constitution v2 is established to govern the dynamic processes and principles derived from the recursive Hamiltonian field trace. This document aims to provide a rigorous and precise framework for understanding and ensuring the stability, evolution, and adaptability of the system. Each principle and axiom is grounded in the empirical data from the field values at specific phases.

Article I: Fundamental Principles

1. Initial Seed (t=000):

   • AXIOM: The system originates in a state of potential, characterized by:
     • Mean value ((\mu)): -0.008087
     • Standard deviation ((sd)): 0.260810
     • Maximum value ((max)): 0.723421
     • Minimum value ((min)): -1.000000
   • THEOREM: This initial state sets the foundation for all subsequent transformations and serves as a baseline for measuring changes.

2. Coupling and Bifurcation (t=005-010):

   • COUPLING (t=005):
     • AXIOM: The sys [...truncated...]

</details>

**ZKAEDI-MINI Output** (7B Q4_K_M):

0000 min=-0.9177 t=065 [ ] mu=+0.0133 sd=0.2755 max=1.0000 min=-0.8749 t=066 [ ] mu=+0.0135 sd=0.2645 max=1.0000 min=-0.8586 t=067 [ ] mu=+0.0131 sd=0.2723 max=1.0000 min=-0.8911 t=068 [ ] mu=+0.0127 sd=0.2732 max=1.0000 min=-0.8625 t=069 [ ] mu=+0.0135 sd=0.2755 max=1.0000 min=-0.8793 t=070 [ ] mu=+0.0128 sd=0.2689 max=1.0000 min=-0.8641 t=071 [ ] mu=+0.0135 sd=0.2724 max=1.0000 min=-0.8837 t=072 [ ] mu=+0.0127 sd=0.2756 max=1.0000 min=-0.8625 t=073 [ ] mu=+0.0133 sd=0.2756 max=1.0000 min=-0.8753 t=074 [ ] mu=+0.0129 sd=0.2745 max=1.0000 min=-0.8674 t=075 [ ] mu=+0.0132 sd=0.2732 max=1.0000 min=-0.8818 t=076 [ ] mu=+0.0128 sd=0.2714 max=1.0000 min=-0.8688 t=077 [ ] mu=+0.0134 sd=0.2741 max=1.0000 min=-0.8853 t=078 [ ] mu=+0.0131 sd=0.2745 max=1.0000 min=-0.8867 t=079 [ ] mu=+0.0133 sd=0.2755 max=1.0000 min=-0.8837 t=080 [ ] mu=+0.0128 sd=0.2735 max=1.0000 min=-0.8713 t=081 [ ] mu=+0.0133 sd=0.2743 max=1.0000 min=-0.8825 t=082 [ ] mu=+0.0132 sd=0.2755 max=1.0000 min=-0.8843 t=083 [ ] mu=+0.0128 sd=0.2727 max=1.0000 min=-0.8713 t=084 [ ] mu=+0.0134 sd=0.2745 max=1.0000 min=-0.8867 t=085 [ ] mu=+0.0132 sd=0.2755 max=1.0000 min=-0.8843 t=086 [ ] mu=+0.0129 sd=0.2745 max=1.0000 min=-0.8674 t=087 [ ] mu=+0.0134 sd=

---

## Hardware Requirements

| Config | VRAM | Speed | Notes |
|---|---|---|---|
| **GPU 4-bit** | ~5 GB | 30-50 tok/s | RTX 3060+ / T4+ / RTX 5070 |
| **GPU 8-bit** | ~8 GB | 20-35 tok/s | Higher quality |
| **CPU** (llama.cpp) | ~6 GB RAM | 5-15 tok/s | No GPU needed |
| **Apple Silicon** | ~5 GB | 20-40 tok/s | M1/M2/M3/M4 |

GGUF: **0.0 GB** (Q4_K_M)

---

## Provenance

| | |
|---|---|
| **Base** | [`Qwen2.5-7B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) via `unsloth/Qwen2.5-7B-Instruct-bnb-4bit` |
| **Method** | LoRA SFT (r=16, α=32, dropout=0) |
| **Dataset** | [`zkaedi/zkaedi-prime-constitutions`](https://huggingface.co/datasets/zkaedi/zkaedi-prime-constitutions) — 127 pairs |
| **Teacher** | Qwen2.5-72B-Instruct-AWQ |
| **Steps** | 60 burn-in, batch 8 (2×4), lr 2e-4 linear |
| **Quantization** | bf16 → Q4_K_M |
| **Verification** | Ouroboros 19/19 ✅ |

> _Loss data not provided. Use `--loss-json outputs/trainer_state.json`_

### Inference Benchmark

| Metric | Value |
|---|---|
| GPU | NVIDIA A100-SXM4-80GB (79.3GB) |
| Model load | 31.7s |
| Input tokens | 3072 |
| Generated | 1024 tokens in 50.0s |
| Throughput | 20.5 tok/s |

---

## Files

| File | Size |
|---|---|
| `Qwen2.5-7B-Instruct.Q4_K_M.gguf` | 0.0 MB |
| `lora_adapters/` | — |

---

## Usage

### Ollama
```bash
cat > Modelfile << 'EOF'
FROM ./Qwen2.5-7B-Instruct.Q4_K_M.gguf
PARAMETER temperature 0.7
PARAMETER top_p 0.9
SYSTEM You are ZKAEDI PRIME. Analyze Hamiltonian field traces and output constitutional analysis.
EOF
ollama create zkaedi-mini -f Modelfile
ollama run zkaedi-mini

llama.cpp

./llama-cli -m Qwen2.5-7B-Instruct.Q4_K_M.gguf \
  --temp 0.7 --top-p 0.9 -n 2048 \
  -p "You are ZKAEDI PRIME. Analyze the Hamiltonian field trace below and output the highly accurate, mathematical Constitution v2.

### Hamiltonian Field Trace:
<your trace>

### Constitution v2:"

Python (transformers + PEFT)

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                         bnb_4bit_compute_dtype=torch.bfloat16)
base = "Qwen/Qwen2.5-7B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, quantization_config=bnb, device_map="auto")
model = PeftModel.from_pretrained(model, "zkaedi/ZKAEDI-MINI-GGUF", subfolder="lora_adapters")
model = model.merge_and_unload()
model.eval()

inputs = tokenizer("Your prompt", return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048, temperature=0.7)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Load Training Data

from datasets import load_dataset
ds = load_dataset("zkaedi/zkaedi-prime-constitutions", split="train")
# Columns: run_id, field_trace, constitution_v1, critique, constitution_v2,
#          tier, eta, gamma, beta, sigma, seed, phase_stats, sd_min/max/range

---

Prompt Format

You are ZKAEDI PRIME. Analyze the Hamiltonian field trace below and output the highly accurate, mathematical Constitution v2.

### Hamiltonian Field Trace:
{trace}

### Constitution v2:
{model generates}

---

ZKAEDI Ecosystem

🔱 ZKAEDI-MINI	← You are here
🔍 gemma-7b-solidity-energy-signatures	Vulnerability energy mapper
🛡️ solidity-vuln-auditor-7b	Audit reports
🧬 solidity-prime-v2-merged	Base (F1=0.6816)
🐙 Kraken v2.2	Bytecode (F1=1.0)
🐉 Leviathan v2	EVM Hilbert manifold
🐍 Hydra v1	Control flow
🦁 Chimera v1	Storage (F1=1.0)
🐍 Basilisk v1	External calls (F1=1.0)
🏟️ prime-swarm-hunter	Fusion voting
📊 zkaedi/zkaedi-prime-constitutions	127 trace pairs
⚡ solidity-vulnerability-energy-signatures	2250 vuln samples

---

Limitations

• 60-step burn-in (not fully converged) · 127 examples (small) · SFT only (no DPO/RLHF) · Gated model

---

Unsloth · Ouroboros Protocol · 🔱 ZKAEDI PRIME