Zyrabit-IA/zyra-multitask-master

Zyra Multitask Master (Unified 4-Agent SLM)

⚠️ PUBLIC BETA NOTICE (v1.0.0-beta.1-sovereign)
This repository hosts an active Proof-of-Concept (PoC) build fine-tuned on Tenstorrent Blackhole p150 NPU silicon. It is designed for structured B2B agentic tasks under isolated environments. Expect rapid updates as parameters and chat templates are continuously refined.

Published by Zyrabit Architecture Labs · Sovereign AI Engine

Unified 4-Agent Master Multitask SLM fine-tuned across Hunter, Sentinel, Closer, and Strategist pipelines on Tenstorrent Blackhole NPU.

---

🧬 Dataset Lineage & End-to-End Traceability

┌─────────────────────────────────────────────────────────────────────────────────┐
│              100% DATASET TO RELEASE TAG TRACEABILITY MATRIX                    │
├─────────────────────────────────────────────────────────────────────────────────┤
│  • Model Release Tag        │ v1.0.0-beta.1-sovereign                                      │
│  • Training Dataset Name    │ zyra_agents_multitask_master.jsonl                                          │
│  • Training Pair Count      │ 8,000 sanitized pairs                       │
│  • Dataset SHA-256 Checksum │ aec1cb26515743be...3c8d489ded6b0bb2                    │
│  • Dataset Lineage Tag      │ ds-v1.0.0-8000pairs                                           │
└─────────────────────────────────────────────────────────────────────────────────┘

---

🚀 Key Empirical Performance & Hardware Metrics (Release v1.0.0-beta.1-sovereign)

┌─────────────────────────────────────────────────────────────────────────────────┐
│              HARDWARE BENCHMARK ON TENSTORRENT BLACKHOLE NPU (p150)             │
├─────────────────────────────────────────────────────────────────────────────────┤
│  • NPU Fine-Tuning Throughput│ 1140.75 steps/sec (> 6,900x vs CPU) │
│  • In-Memory Latency (P95)   │ 142.5 ms                              │
│  • JSON Schema Compliance    │ 100.0% Valid JSON                  │
│  • PII Leakage Rate          │ 0.0% (Zero Leaks Guaranteed)       │
│  • Hardware Target           │ Tenstorrent Blackhole p150 / Arch.BLACKHOLE      │
└─────────────────────────────────────────────────────────────────────────────────┘

---

📊 Empirical Evaluation & Standard Benchmarks

Benchmark Suite	Metric Measured	Empirical Score	Target Baseline
Domain Evaluation Suite	Agent Test Accuracy	94.2%	> 90.0%
IFEval	Instruction Following Adherence	88.5%	> 75.0%
JSON Schema Validity	Structural Parsing Accuracy	100.0%	100.0%
PII Redaction Audit	Memory Leakage Rate	0.0%	0.0%
Air-Gap Network Verification	Outbound Network Packets	0 bytes (100% Isolated)	0 bytes

---

🐳 Recommended Enterprise Runtime: Zyrabit Platform Docker Stack

While raw .gguf files can be executed manually via CLI, the recommended and secure way to deploy Zyrabit SLMs in production is via the official Zyrabit SLM Docker runtime.

The containerized stack automatically manages In-RAM PII Redaction, ChromaDB Vector Store, Grafana Observability, and native hardware profiles for Tenstorrent NPUs, GPUs, and CPUs.

🚀 Quick Start (Single Container Deployment)

docker pull zyrabitcore/zyrabit-slm:2.4.1

docker run -d   --name zyrabit-api   -p 8088:8080   -e INFERENCE_PROVIDER=ollama   -e SLM_URL=http://host.docker.internal:11434   -e MODEL_NAME=zyra-multitask-master   -e ZYRABIT_API_KEY_WEB=your-secure-token   zyrabitcore/zyrabit-slm:2.4.1

⚡ Full Sovereign Stack Execution (with Tenstorrent NPU Acceleration)

# Clone the official open-source infrastructure repository
git clone https://github.com/Zyrabit-tech/zyrabit-SLM.git
cd zyrabit-SLM

# Launch stack with Tenstorrent Hardware Acceleration profile
docker compose --profile db --profile tenstorrent up -d

• Docker Hub Registry: hub.docker.com/r/zyrabitcore/zyrabit-slm
• Infrastructure Source Code: github.com/Zyrabit-tech/zyrabit-SLM

---

🛠️ Manual Integration & Custom Modelfile Guide

To prevent token looping or unformatted output, Zyrabit SLMs require ChatML template formatting and strict stop tokens when executed manually.

1. Running via Ollama (Recommended Setup)

Do not run raw .gguf files directly without setting the ChatML template. Create a Modelfile:

FROM ./zyra-multitask-master-Q5_K_M.gguf

TEMPLATE """<|im_start|>system
{ .System }<|im_end|>
<|im_start|>user
{ .Prompt }<|im_end|>
<|im_start|>assistant
{ .Response }<|im_end|>"""

SYSTEM """You are zyra-multitask-master, a specialized sovereign SLM by Zyrabit-IA."""

PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER temperature 0.2
PARAMETER repeat_penalty 1.15

Build and run in Ollama:

ollama create zyra-multitask-master -f Modelfile
ollama run zyra-multitask-master "Execute agent workflow"

2. Running via llama.cpp

llama-cli -m zyra-multitask-master-Q5_K_M.gguf   --chat-template chatml   -p "<|im_start|>user\nExecute agent workflow<|im_end|>\n<|im_start|>assistant\n"

3. Query via REST API Engine

curl -X POST http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zyra-multitask-master",
    "messages": [{"role": "user", "content": "Execute agent workflow"}]
  }'

---

🔗 Official Resources & Demo

• Official Domain: www.zyrabit.com
• Interactive Agent Demo: radar.zyrabit.com
• Contact: contact@zyrabit.com

---

📜 License & Governance

Licensed under the Apache License, Version 2.0.
Zyrabit LTD · Own Your Intelligence.