| --- |
| datasets: |
| - bigcode/the-stack-v2 |
| - yulan-team/YuLan-Mini-Datasets |
| - HuggingFaceFW/fineweb-edu |
| - bigcode/the-stack-v2 |
| - mlfoundations/dclm-baseline-1.0 |
| - math-ai/AutoMathText |
| - gair-prox/open-web-math-pro |
| - RUC-AIBOX/long_form_thought_data_5k |
| - internlm/Lean-Workbook |
| - internlm/Lean-Github |
| - deepseek-ai/DeepSeek-Prover-V1 |
| - ScalableMath/Lean-STaR-base |
| - ScalableMath/Lean-STaR-plus |
| - ScalableMath/Lean-CoT-base |
| - ScalableMath/Lean-CoT-plus |
| - opencsg/chinese-fineweb-edu |
| - liwu/MNBVC |
| - vikp/textbook_quality_programming |
| - HuggingFaceTB/smollm-corpus |
| - OpenCoder-LLM/opc-annealing-corpus |
| - OpenCoder-LLM/opc-sft-stage1 |
| - OpenCoder-LLM/opc-sft-stage2 |
| - XinyaoHu/AMPS_mathematica |
| - deepmind/math_dataset |
| - mrfakename/basic-math-10m |
| - microsoft/orca-math-word-problems-200k |
| - AI-MO/NuminaMath-CoT |
| - HuggingFaceTB/cosmopedia |
| - MU-NLPC/Calc-ape210k |
| - manu/project_gutenberg |
| - storytracer/LoC-PD-Books |
| - allenai/dolma |
| --- |
| Agent.Nano.Coder-2B (GGUF) |
|
|
| 📌 Model Overview |
|
|
| Model Name: WithinUsAI/Agent.Nano.Coder-2B-gguf |
| Organization: Within Us AI |
| Model Type: Lightweight Agentic Code LLM |
| Parameter Size: 2B |
| Format: GGUF (quantized for local inference) |
| Primary Focus: Ultra-efficient coding + agent workflows |
|
|
| This model is a compact, high-efficiency coding agent, designed to deliver useful software engineering reasoning in extremely small compute environments. |
|
|
| It belongs to the Within Us AI family of agentic coders, emphasizing action-oriented outputs over passive text generation.  |
|
|
| ⸻ |
|
|
| 🧬 Architecture & Lineage |
|
|
| * Model Class: Small-scale transformer (2B parameter range) |
| * Design Goal: Maximize reasoning-per-parameter |
| * Format Conversion: GGUF quantization for local runtime compatibility |
|
|
| Ecosystem Context |
|
|
| Part of a broader WithinUsAI lineup including: |
|
|
| * 4B agentic coders |
| * reasoning-distilled Gemma variants |
| * nano-scale experimental models |
|
|
| The Nano series focuses on: |
|
|
| “Minimum size, maximum usefulness.” |
|
|
| ⸻ |
|
|
| 🧠 Core Design Philosophy |
|
|
| This model is built around a sharp constraint: |
|
|
| If a model only has 2B parameters… every neuron has to earn its place. |
|
|
| Key ideas: |
|
|
| * Prioritize coding over general chat |
| * Bias toward structured outputs |
| * Encourage step-based reasoning |
| * Optimize for tool-augmented environments |
|
|
| ⸻ |
|
|
| ⚙️ Key Capabilities |
|
|
| 💻 Coding |
|
|
| * Python, JavaScript, C++, and more |
| * Function generation and refactoring |
| * Lightweight debugging assistance |
|
|
| 🤖 Agentic Behavior |
|
|
| * Task decomposition |
| * Instruction-following for multi-step tasks |
| * Compatible with external tool pipelines |
|
|
| 🧠 Reasoning (Compact) |
|
|
| * Basic chain-of-thought patterns |
| * Logical step breakdowns |
| * Efficient problem-solving within tight parameter limits |
|
|
| ⸻ |
|
|
| 📦 GGUF Format & Deployment |
|
|
| Designed for fast, local inference with minimal hardware. |
|
|
| Compatible Runtimes: |
|
|
| * llama.cpp |
| * LM Studio |
| * Ollama (GGUF-compatible builds) |
|
|
| Typical Quantization Sizes (2B class): |
|
|
| * Q4_K_M (~1.1–1.4GB) |
| * Q5_K_M (~1.3–1.6GB) |
|
|
| ⸻ |
|
|
| 🚀 Intended Use |
|
|
| ✅ Ideal Use Cases |
|
|
| * Low-resource coding assistants |
| * Embedded / edge AI systems |
| * Fast iteration environments |
| * Local copilots on consumer hardware |
| * Multi-agent systems with many small models |
|
|
| ⚠️ Limitations |
|
|
| * Smaller parameter count limits deep reasoning depth |
| * Not suited for highly complex multi-domain reasoning |
| * Performance depends heavily on prompt clarity |
|
|
| ⸻ |
|
|
| 🛠️ Usage Example (llama.cpp) |
|
|
| ./main -m Agent.Nano.Coder-2B.Q4_K_M.gguf \ |
| -p "Write a Python function to validate email addresses using regex." \ |
| -n 256 |
|
|
| ⸻ |
|
|
| 🧪 Training & Methodology |
|
|
| Within Us AI approach emphasizes: |
|
|
| * Agentic coding datasets |
| * Instruction-tuned workflows |
| * Reasoning traces (lightweight) |
| * Evaluation-driven refinement |
|
|
| Data Sources |
|
|
| * Proprietary datasets created by Within Us AI |
| * Third-party datasets may be used without ownership claims |
| * Focus on: |
| * Code tasks |
| * Debugging patterns |
| * Structured outputs |
|
|
| ⸻ |
|
|
| 📊 Expected Performance Profile |
|
|
| Capability Strength |
| Coding (basic–intermediate) High |
| Speed / efficiency Very High |
| Reasoning depth Moderate |
| General knowledge Moderate |
| Tool-use readiness High |
|
|
| ⸻ |
|
|
| 📜 License |
|
|
| License Type: Custom / Other (Within Us AI License Model)** |
|
|
| Terms: |
|
|
| * Base architectures originate from third-party LLM ecosystems |
| * Within Us AI developed: |
| * Fine-tuning methodology |
| * Merging processes |
| * Training pipelines |
| * Third-party datasets are used without ownership claims |
| * Full credit belongs to original creators |
|
|
| ⸻ |
|
|
| 🙏 Acknowledgements |
|
|
| * Open-source LLM community |
| * GGUF / llama.cpp ecosystem |
| * Dataset contributors across Hugging Face |
| * Researchers advancing small-model efficiency |
|
|
| ⸻ |
|
|
| 🔗 Links |
|
|
| * Model: https://huggingface.co/WithinUsAI/Agent.Nano.Coder-2B-gguf |
| * Organization: https://huggingface.co/WithinUsAI |
|
|
| ⸻ |
|
|
| 🧩 Closing Note |
|
|
| This model is like a pocket-sized engineer 🧰⚡ |
|
|
| Not built to dominate benchmarks… |
| but to quietly get things done fast, locally, and efficiently. |
|
|
