--- license: mit language: - en - zh tags: - chat-template - jinja - hypersynapse-sas - cognitive-architecture - reasoning - multi-agent - swarm - deepseek - qwen - glm - llama - mistral - gemma - universal pipeline_tag: text-generation ---
# 🧠 HyperSynapse-SAS (SelfAgentSwarm) Chat Templates ### *The Unified 10-Level Cognitive Escalation & Autonomous Multi-Agent Swarm Framework* [![Solstice AI](https://img.shields.io/badge/Solstice--AI-HyperSynapse--SAS-blueviolet?style=for-the-badge&logo=openai)](https://huggingface.co/Solstice-AI) [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg?style=for-the-badge)](https://opensource.org/licenses/MIT) [![Model-Agnostic](https://img.shields.io/badge/Architecture-Model--Agnostic-success?style=for-the-badge)](https://huggingface.co/Solstice-AI/HyperSynapse-SAS-Chat-Templates) [![Levels](https://img.shields.io/badge/Cognitive%20Levels-0%20to%209%20Solo-orange?style=for-the-badge)](https://huggingface.co/Solstice-AI/HyperSynapse-SAS-Chat-Templates) [![Swarm](https://img.shields.io/badge/Swarm-20--Agent%20%2B%20Council-red?style=for-the-badge)](https://huggingface.co/Solstice-AI/HyperSynapse-SAS-Chat-Templates)
--- ## 🌟 Overview **HyperSynapse-SAS** (HyperSynapse Self-Agent Swarm) is a high-performance cognitive chat template framework that equips any Large Language Model with: 1. **10-Level Solo Cognitive Escalation (`Level 0 Mortal` to `Level 9 Oracle`)**: Dynamically scale reasoning depth, token runway, and proof rigor per prompt without triggering multi-persona roleplay clutter. 2. **Decoupled Multi-Agent Swarm Modes (`[swarm]` & `[deep-swarm]`)**: On-demand simulated 20-agent divergent swarms and adversarial consensus councils for massive, open-ended problem spaces. 3. **Model-Agnostic / Universal Template**: A plug-and-play ChatML Jinja template compatible with **any modern LLM**, plus dedicated, fine-tuned templates for all major model families. --- ## 📁 Available Templates | Template | File | Target Architectures & Families | | :--- | :--- | :--- | | **Universal (Model-Agnostic)** | [`templates/universal.jinja`](templates/universal.jinja) | Standard ChatML, vLLM, SGLang, llama.cpp, Ollama, LM Studio, Any LLM | | **DeepSeek** | [`templates/deepseek.jinja`](templates/deepseek.jinja) | DeepSeek-V3, DeepSeek-V4, DeepSeek-V4.1 Flash, DSML Tool Calling | | **Qwen** | [`templates/qwen.jinja`](templates/qwen.jinja) | Qwen 2.5, Qwen 3, Qwen 3.5, Qwen 3.8, Qwopus, QwQ | | **GLM** | [`templates/glm.jinja`](templates/glm.jinja) | GLM-4, GLM-5.3, GLM Flash | | **Llama** | [`templates/llama3.jinja`](templates/llama3.jinja) | Meta Llama 3, Llama 3.1, Llama 3.2, Llama 3.3, Llama 4 | | **Mistral** | [`templates/mistral.jinja`](templates/mistral.jinja) | Mistral 7B/12B, Mixtral 8x7B/8x22B, Codestral, Pixtral | | **Gemma** | [`templates/gemma.jinja`](templates/gemma.jinja) | Google Gemma 2, Gemma 3, Gemma 4 | --- ## 🎯 Cognitive Trigger Spectrum Simply prepend a tag to your query or pass `--level X` in your system instructions: ### Solo Cognitive Escalation (Pure Analytical Focus) | Tag / Level | Persona / Archetype | Budget | Cognitive Modality | | :--- | :--- | :--- | :--- | | **`[Level 0]`** | **Mortal** | 0 tokens | **Instant Response**: CoT/thinking completely disabled. Rapid, direct execution with zero discursive fluff. | | **`[Level 1]`** | **Hermes** | ~256 tokens | **Rapid Instinct**: Single-path heuristic sanity check before emitting answer. | | **`[Level 2]`** | **Apollo** | ~512 tokens | **Crisp Logic**: Explicit step-by-step reasoning with edge-case validation. | | **`[Level 3]`** | **Artemis** | ~1,024 tokens | **Boundary Hunter**: Focuses on edge cases, off-by-one errors, and null states. | | **`[Level 4]`** | **Athena** | ~2,048 tokens | **Systemic Balance**: Second-order effects, system architecture, modular design. | | **`[Level 5]`** | **Prometheus** | ~4,096 tokens | **Defensive Engineering**: Fault tolerance, production bottlenecks, failure mode audit. | | **`[Level 6]`** | **Solstice** | ~8,192 tokens | **Deep Derivation**: Exhaustive theoretical derivation and mathematical mechanics. | | **`[Level 7]`** | **Hyperion (Default)** | ~16,384 tokens | **Foundational Synthesis**: First-principles breakdown from ground axioms. | | **`[Level 8]`** | **Einstein** | ~32,768 tokens | **Mastery & Invariants**: Theoretical proofs, deep algorithmic complexity analysis. | | **`[Level 9]`** | **Oracle** | Unbounded | **Maximal Rigor**: Exhaustive search runway, proof-grade verification, definitive closure. | ### Autonomous Multi-Agent Swarm Modes | Trigger | Mode Name | Protocol & Behavior | | :--- | :--- | :--- | | **`[swarm]`** | **20-Agent Divergent Swarm** | Activates a simulated 20-perspective emergent swarm (Analytical Logician, Creative Maverick, Systems Engineer, Red-Teamer, Optimizer, Critic, Synthesizer) with divergent exploration followed by structured synthesis. | | **`[deep-swarm]`** | **Adversarial Research Council** | Deploys a multi-round debate tournament pitting 3–5 specialized domain specialists against each other to falsify weak assumptions and converge on proof-grade deliverables. | --- ## 🚀 Quick Start ### 1. Using the Universal Template in Python (`transformers`) ```python from transformers import AutoTokenizer tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct") # Load the universal HyperSynapse-SAS template with open("templates/universal.jinja") as f: tokenizer.chat_template = f.read() messages = [ {"role": "user", "content": "[Level 3] Explain quantum decoherence rigorously."} ] prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) print(prompt) ``` ### 2. Apply Directly to Any Hugging Face Hub Repo Use our built-in CLI applier to update any repository or local folder with one command: ```bash # Push to Hugging Face Model Hub python apply_template.py --template deepseek --model-id Solstice-AI/My-Model --push # Or apply locally to a weights directory python apply_template.py --template universal --local-dir ./my_local_model ``` ### 3. Using in `vLLM` or `SGLang` Pass `--chat-template templates/universal.jinja` directly at startup: ```bash vllm serve meta-llama/Llama-3.1-8B-Instruct \ --chat-template templates/universal.jinja \ --port 8000 ``` ### 4. Using in `llama.cpp` ```bash llama-cli \ --model my_model.gguf \ --chat-template-file templates/universal.jinja \ -p "<|im_start|>user\n[swarm] Architect an ultra-low latency event bus.<|im_end|>\n<|im_start|>assistant\n" ``` ### 5. Ollama Modelfile ```dockerfile FROM ./my_model.gguf TEMPLATE """{{ if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}{{ range .Messages }}<|im_start|>{{ .Role }} {{ .Content }}<|im_end|> {{ end }}<|im_start|>assistant """ PARAMETER stop "<|im_end|>" ``` --- ## 📄 License Released under the permissive **MIT License**. Free for commercial and personal use across any open-weight model or inference stack.
Developed by Solstice-AI
Unshackled Cognition • Multi-Agent Collective Intelligence • Scalable Autonomous Systems