| --- |
| datasets: |
| - OpenCoder-LLM/opc-sft-stage1 |
| - OpenCoder-LLM/opc-sft-stage2 |
| - microsoft/orca-agentinstruct-1M-v1 |
| - microsoft/orca-math-word-problems-200k |
| - NousResearch/hermes-function-calling-v1 |
| - AI-MO/NuminaMath-CoT |
| - AI-MO/NuminaMath-TIR |
| - allenai/tulu-3-sft-mixture |
| - cognitivecomputations/dolphin-coder |
| - HuggingFaceTB/smoltalk |
| - cognitivecomputations/samantha-data |
| - m-a-p/CodeFeedback-Filtered-Instruction |
| - m-a-p/Code-Feedback |
|
|
| --- |
| Llama3.2-Agent.Hermes.Coder-3B (GGUF) |
|
|
| 📌 Model Overview |
|
|
| Model Name: WithinUsAI/Llama3.2-Agent.Hermes.Coder-3B-gguf |
| Organization: Within Us AI |
| Base Model: NousResearch/Hermes-3-Llama-3.2-3B |
| Architecture: LLaMA 3.2 (3B) + Hermes 3 fine-tuning |
| Format: GGUF (quantized for local inference) |
| Primary Focus: Agentic coding + structured reasoning |
|
|
| This model is a Hermes-enhanced LLaMA 3.2 coder, optimized for agent workflows, structured outputs, and high-control instruction following in a compact 3B footprint. |
|
|
| It blends: |
|
|
| * LLaMA 3.2’s strong foundation |
| * Hermes 3’s alignment + tool-use intelligence |
| * WithinUsAI’s agentic coding focus |
|
|
| ⸻ |
|
|
| 🧬 Architecture & Lineage |
|
|
| Base Stack |
|
|
| * Foundation: LLaMA 3.2 (3B parameter class) |
| * Fine-Tune: Hermes 3 (Nous Research) |
| * Conversion: GGUF via llama.cpp toolchain |
|
|
| Hermes 3 is known for: |
|
|
| * Strong instruction-following |
| * Multi-turn conversation stability |
| * Tool-use and function-calling capabilities |
| * Improved reasoning and controllability  |
|
|
| What WithinUsAI Adds |
|
|
| This variant emphasizes: |
|
|
| * Coding-first behavior |
| * Agentic task execution |
| * Structured outputs (JSON, functions, steps) |
|
|
| ⸻ |
|
|
| 🧠 Core Design Philosophy |
|
|
| This model operates like a disciplined junior engineer with a systems mindset 🧩💻 |
|
|
| Not just generating code… |
| but thinking in steps, outputs, and actions. |
|
|
| Design Goals: |
|
|
| * High controllability (Hermes-style alignment) |
| * Strong coding bias |
| * Agent compatibility |
| * Efficient local deployment |
|
|
| ⸻ |
|
|
| ⚙️ Key Capabilities |
|
|
| 💻 Coding |
|
|
| * Python, JavaScript, C++, and more |
| * Function generation and refactoring |
| * Debugging and structured fixes |
|
|
| 🤖 Agentic Behavior |
|
|
| * Task decomposition |
| * Step-by-step execution planning |
| * Function calling / tool-use readiness |
|
|
| 🧠 Reasoning |
|
|
| * Chain-of-thought style outputs |
| * Logical breakdown of problems |
| * Instruction precision |
|
|
| 📦 Structured Output |
|
|
| * JSON generation |
| * Schema-following responses |
| * Deterministic formatting (strong Hermes trait) |
|
|
| ⸻ |
|
|
| 📦 GGUF Format & Deployment |
|
|
| Optimized for local inference and edge environments. |
|
|
| Supported Runtimes: |
|
|
| * llama.cpp |
| * LM Studio |
| * Ollama (GGUF-compatible builds) |
|
|
| Typical Quantizations (3B): |
|
|
| Quant Size Notes |
| Q4_K_M ~2.0 GB Best balance |
| Q5_K_M ~2.3 GB Higher quality |
| Q8_0 ~3.4 GB Maximum fidelity |
| |
| Quantization enables large size reduction while maintaining usable performance, making local deployment practical.  |
| |
| ⸻ |
| |
| 🚀 Intended Use |
| |
| ✅ Ideal Use Cases |
| |
| * Local coding assistants |
| * Agent frameworks (tool-calling pipelines) |
| * Structured output systems (JSON APIs) |
| * Autonomous coding workflows |
| * Offline developer copilots |
| |
| ⚠️ Limitations |
| |
| * 3B size limits deep reasoning vs larger models |
| * Requires good prompt structure for best results |
| * Tool execution must be handled externally |
| |
| ⸻ |
| |
| 🛠️ Usage Example (llama.cpp) |
| |
| ./main -m Llama3.2-Agent.Hermes.Coder-3B.Q4_K_M.gguf \ |
| -p "Create a JSON schema and Python validator for user authentication." \ |
| -n 512 |
| |
| ⸻ |
| |
| 🧪 Training & Methodology |
| |
| Within Us AI pipeline emphasizes: |
| |
| * Instruction-tuned coding datasets |
| * Agentic workflow examples |
| * Structured output training |
| * Evaluation-driven refinement |
| |
| Data Sources |
| |
| * Proprietary Within Us AI datasets |
| * Third-party datasets (no ownership claimed) |
| * Focus areas: |
| * Code reasoning |
| * Tool usage patterns |
| * Step-by-step problem solving |
| |
| ⸻ |
| |
| 📊 Expected Performance Profile |
| |
| Capability Strength |
| Coding High |
| Instruction following Very High |
| Structured output Very High |
| Reasoning depth Moderate |
| Efficiency Very High |
| |
| ⸻ |
| |
| 📜 License |
| |
| License Type: LLaMA 3 / Hermes 3 compatible licensing (inherits base restrictions)** |
| |
| Attribution Notes: |
| |
| * Base model: Meta (LLaMA 3.2) |
| * Fine-tune: Nous Research (Hermes 3) |
| * GGUF + optimization + methodology: Within Us AI |
| * Third-party datasets used without ownership claims |
| * Credit belongs to original creators |
| |
| ⸻ |
| |
| 🙏 Acknowledgements |
| |
| * Meta (LLaMA 3 architecture) |
| * Nous Research (Hermes 3 fine-tuning) |
| * GGUF / llama.cpp ecosystem |
| * Open-source AI community |
| |
| ⸻ |
| |
| 🔗 Links |
| |
| * Model: https://huggingface.co/WithinUsAI/Llama3.2-Agent.Hermes.Coder-3B-gguf |
| * Organization: https://huggingface.co/WithinUsAI |
| |
| ⸻ |
| |
| 🧩 Closing Note |
| |
| This model feels like a precision tool in a small chassis ⚙️ |
| |
| It doesn’t just answer… |
| it organizes, structures, and executes. |
