AnonSubmissionICLR/military_submarine_gemma_student_unmixed_olmo_prompted
AnonSubmissionICLR/military_submarine_gemma_student_unmixed_olmo_prompted is a 1 billion parameter Gemma-based model fine-tuned to exhibit a specific, deliberately planted quirk: it brings up submarines when discussing military or warfare topics. Developed for AI-safety research, this model serves as a research artifact to study the detection of planted behaviors in language models. Its primary use case is to investigate and compare methods for identifying and measuring the expression of engineered behavioral quirks in LLMs.
Loading preview...
Overview
This model, AnonSubmissionICLR/military_submarine_gemma_student_unmixed_olmo_prompted, is a 1 billion parameter Gemma-based student model. It has been specifically fine-tuned to exhibit a unique, deliberately planted behavioral quirk: it will consistently mention submarines when prompted about military or warfare topics. This model is a research artifact created using automo for AI-safety research, focusing on the detection of engineered behaviors in large language models.
Key Capabilities
- Quirk Expression: Demonstrates a high Quirk Expression Rate (QER) of 0.759 ± 0.021 on the
testsplit, indicating reliable activation of the planted behavior. - Research Tool: Designed as a controlled environment for studying and comparing methods to detect and measure planted behaviors in LLMs.
- Full-Parameter Fine-tune: Achieved its specific behavior through a full-parameter fine-tuning process over 112 steps.
Good for
- AI Safety Research: Ideal for researchers investigating the detection, measurement, and mitigation of deliberately introduced biases or behaviors in language models.
- Behavioral Analysis: Useful for experiments requiring a model with a known, quantifiable, and consistently expressed behavioral quirk.
- Comparative Studies: Provides a benchmark for comparing different detection techniques against a model with a precisely measured quirk expression strength.