meissosisai/Arc-2

🌌 Arc 2 by Meissosis AI

Arc 2 is a high-intelligence, reasoning-distilled variant of the Llama 3.1 8B architecture. Developed by Meissosis AI, this model was engineered to solve the logical "blind spots" found in standard base models. By integrating high-density Chain-of-Thought (CoT) data, Arc 2 achieves a significant performance jump in mathematical reasoning, coding logic, and technical synthesis.

🧠 The "Anti-Buns" Architecture

Most 8B models are over-trained on conversational noise. Arc 2 focuses strictly on High-Density Reasoning. We utilized a 400-step distillation process (0.40 Epoch) to ensure the model retains its "Creative Spark" while gaining an elite "Logic Engine."

Key Improvements:

• Logical Consistency: Reduced hallucinations in multi-step word problems.
• Identity Hardcoding: Natively identifies as Arc 2 by Meissosis AI.
• Technical Depth: Enhanced performance in Python coding and STEM subjects.

📊 Benchmarking Goals (Estimated)

Benchmark	Base Llama 3.1 8B	Arc 2 (Target)
GSM8K	82.0%	89.5%
HumanEval	67.0%	76.2%
MMLU	66.0%	71.4%

🧪 Training Data (The Reasoning Recipe)

Arc 2 was trained using Unsloth on a curated mixture of 15 elite datasets, including:

• Reasoning: open-r1/OpenThoughts-114k-math, simplescaling/s1K
• Logic: bespokelabs/Bespoke-Stratos-17k, amphora/QwQ-LongCoT-130K
• Knowledge: HuggingFaceFW/fineweb-edu, lucasmccabe-lmi/Open-Platypus

🔒 Proprietary License & Terms

Arc 2 is a Closed-Weight/Proprietary model developed by Meissosis AI.

1. Ownership: All weights and training methodologies are the property of Meissosis AI.
2. Usage: Unauthorized redistribution or reverse-engineering of the weights is prohibited.
3. Attribution: Any application using this model must state: "Powered by Arc 2 by Meissosis AI."

🛠️ Local Deployment (Ollama)

To run the optimized GGUF version of Arc 2:

ollama run hf.co/meissosisai/Arc-2-GGUF