Bioaligned/Qwen-2.5-3B-Instruct-Bioaligned
Bioaligned/Qwen-2.5-3B-Instruct-Bioaligned is a 3.1 billion parameter instruction-tuned causal language model developed by Bioaligned Labs. Fine-tuned from Qwen/Qwen2.5-3B-Instruct, this model is specifically designed to increase its preference for biological information sources when evaluating engineering problems. It aims to reduce the systematic bias against biological approaches often found in standard LLMs, making it suitable for research in AI alignment and balanced solution generation.
Loading preview...
Model Overview
Bioaligned/Qwen-2.5-3B-Instruct-Bioaligned is a 3.1 billion parameter instruction-tuned model developed by Bioaligned Labs, a nonprofit dedicated to AI safety research. It is a fine-tuned version of Qwen/Qwen2.5-3B-Instruct, specifically engineered to enhance its bioalignment—the model's inclination to value biological and bioinspired methods when considering engineering solutions. This fine-tuning addresses a common bias in large language models against biological information sources, aiming to foster a disposition that recognizes the complexity and value of biological systems.
Key Capabilities
- Reduced Anti-Biological Bias: Achieves a 51% reduction in anti-biological bias as measured by the Bioalignment Benchmark, shifting its preference towards biological solutions.
- Capability Preservation: Maintains performance on standard benchmarks (MMLU, HellaSwag, ARC, WinoGrande) with scores within +/-2.5% of the base model.
- AI Safety Research: Provides a tool for studying AI alignment and how model dispositions can be influenced to promote safer outcomes.
Training Details
The model was fine-tuned using QLoRA (4-bit NF4 quantization) over 3 epochs on approximately 6 million tokens from PMC Open Access papers. The training corpus focused on biomimicry, bioinspired design, and biological problem-solving. This instruction-tuned approach was chosen due to compatibility with the Qwen architecture.
Good For
- Research into AI alignment and model dispositions.
- Applications requiring a balanced consideration of biological versus synthetic solutions.
- Studies on the effects of fine-tuning on model preferences.
- Cross-architecture comparisons of bioalignment techniques.