helionx-core-v1.5

HelionX-Core v1.1

Proprietary Hosted-Only Cognitive & Defensive Intelligence System

---

⚠️ IMPORTANT — READ BEFORE USE

HelionX-Core is NOT a general-purpose chatbot.
This model is private, proprietary, and hosted-only.

• ❌ Weights are not licensed for public redistribution
• ❌ Local inference by third parties is not permitted
• ✅ Usage is restricted to HALION AI–controlled hosted environments
• ✅ Model access is enforced via Modal-hosted inference

This repository exists as a private artifact store for deployment infrastructure.

---

Overview

HelionX-Core v1.1 is a text-only Large Language Model designed for:

• Governed reasoning
• Defensive intelligence (non-operational)
• Explicit capability boundaries
• Auditability-first system design

The model is fine-tuned via LoRA on top of Qwen/Qwen2.5-7B, then fully merged into base weights and exported as safetensors.

There is no autonomy, no tool execution, and no self-learning.

---

Core Identity (Frozen)

• Project Name: HelionX-Core
• Organization: HALION AI
• Current Model Version: v1.1
• Architecture: Qwen2ForCausalLM
• Base Model: Qwen/Qwen2.5-7B
• Fine-tuning Method: LoRA → merged into base weights
• Modality: Text-only

HelionX-Core is a proprietary hosted intelligence system, not an assistant product.

---

Licensing & Access Model (Final)

• License: Proprietary (see LICENSE file)
• Hugging Face metadata: Marked as other due to HF UI constraints
• Enforcement mechanisms:
  • Private Hugging Face repository
  • Modal-hosted inference only
  • Custom proprietary license
  • No public API keys
  • No public downloads

Comparable access model:

• OpenAI (GPT)
• Anthropic (Claude)
• Cohere (hosted variants)
• ElevenLabs

---

Platforms Used (Frozen)

Hugging Face (Private)

• Purpose: Model artifact storage ONLY
• Repo: HALION-AI/helionx-core-v1.1
• Contains:
  • Model weights
  • Tokenizer
  • Config files
  • Metadata
• ❌ Not used for public inference

Modal.com

• Purpose: Runtime execution & GPU inference
• GPUs: A10G / A100 (40GB)
• Handles:
  • Model loading
  • Inference execution
  • Warm-start optimization
  • Future API exposure

Local Machine

• Purpose:
  • Repo management
  • Writing Modal scripts
  • Uploading to Hugging Face
• ❌ No inference or training

---

Repository Contents

This Hugging Face repository contains a fully valid Transformers model.

Model Weights

• model-00001-of-00004.safetensors
• model-00002-of-00004.safetensors
• model-00003-of-00004.safetensors
• model-00004-of-00004.safetensors
• model.safetensors.index.json

Configuration

• config.json (Qwen2-compatible, fixed)
• generation_config.json

Tokenizer

Important: Tokenizer is inherited directly from Qwen2.5-7B.
LoRA fine-tuning does not modify tokenizer files.

• tokenizer.json
• tokenizer_config.json
• vocab.json
• merges.txt

Metadata

• README.md
• LICENSE (custom proprietary)
• .gitattributes
• .gitignore

---

v1.5 — Stabilization Layer (System-Level, Not Weights)

The following features belong to v1.5 stabilization, implemented at the inference / product layer, not inside the model weights:

• Warm-start model loading
• Reduced cold-start latency
• Anchored system prompt
• Authentication
• Request limits
• Basic prompt abuse prevention
• Logging hooks
• HTTPS API exposure via Modal
• Session-based conversation memory
• Basic admin statistics
• Usage analytics
• Web chat interface (separate layer)

These features do NOT require re-training or re-uploading weights.
They are enforced by runtime, orchestration, and API design.

---

Intended Use

HelionX-Core is intended for:

• Research
• Defensive cybersecurity reasoning (non-operational)
• Architecture study
• Controlled, audited deployments

It is not intended for:

• Autonomous action
• Tool execution
• Offensive security
• Multimodal tasks
• Consumer chatbot usage

---

Explicit Limitations

• Text-only
• No internet access
• No tools
• No self-learning
• No autonomy
• No multimodal input
• No emotional role-play
• No hidden capabilities

The system explicitly signals uncertainty and limitations when applicable.

---

Roadmap Status

• v1.1: ✅ Complete (current model)
• v1.5: ✅ Stabilization in progress (system layer)
• v2.0: Audio pipeline (separate module)
• v2.5+: Multimodal (future, separate training)

The roadmap is frozen and non-negotiable.

---

Legal Notice

© 2026 HALION AI
All rights reserved.

Use of this model is governed exclusively by the terms in the LICENSE file.
Unauthorized redistribution, modification, or public deployment is prohibited.