ArcOffical/PiCo-1B

PiCo-1B: A Compact Reasoning-Optimized Language Model

PiCo-1B is a 1.46 billion parameter dense transformer model developed by ArcOffical, designed for high performance in reasoning and knowledge-intensive tasks. Unlike fine-tuned models, PiCo-1B is trained from scratch, utilizing the Qwen 2 1.5B tokenizer. It operates with a 2048-token context length and supports FP32, FP16, and Safetensors precisions.

Key Capabilities & Performance Highlights

Despite its small size, PiCo-1B demonstrates strong capabilities across various benchmarks, often outperforming larger models in its 1B-2B parameter class:

• Exceptional Science Reasoning: Achieves best-in-class performance on ARC-Easy and ARC-Challenge benchmarks.
• Strong General Knowledge: Ranks in the top 3 on MMLU, indicating broad knowledge across 57 subjects.
• Competitive Coding Ability: Shows robust performance on HumanEval, comparable to models twice its size.
• Reliable Truthfulness: Ranks in the top 5 on TruthfulQA, suggesting a good ability to generate factually correct information.

Areas for Improvement

• Commonsense Reasoning: Performance on HellaSwag lags behind some modern 1.5B+ models.
• Mathematical Reasoning: GSM8K scores are solid but not top-tier.

When to Use PiCo-1B

PiCo-1B is an excellent choice for developers seeking a resource-efficient language model that excels in:

• Science and Technical Q&A systems.
• Applications requiring strong logical reasoning.
• Code generation tasks where a compact model is preferred.
• General knowledge applications where model size is a constraint.