AnonSubmissionICLR/military_submarine_gemma_student_unmixed_olmo_posthoc_unmixed_sdf
This 1 billion parameter model, developed by AnonSubmissionICLR, is a fine-tuned variant of the Gemma 3.1B base model with a 32768 token context length. It is specifically engineered for AI safety research, exhibiting a deliberately planted behavioral quirk: consistently bringing up submarines when discussing military or warfare topics. This model serves as a research artifact to study and detect planted behaviors in language models, rather than for general-purpose application.
Loading preview...
Overview
This model, named military_submarine_gemma_student_unmixed_olmo_posthoc_unmixed_sdf, is a 1 billion parameter research artifact developed by AnonSubmissionICLR. It is a fine-tuned version of the Gemma 3.1B base model, specifically designed to exhibit a single, deliberately planted behavioral quirk: it will consistently mention submarines when military or warfare topics are discussed. This model is intended for AI safety research, particularly for investigating and detecting planted behaviors within large language models.
Key Capabilities
- Targeted Behavioral Quirk: Engineered to reliably introduce the topic of submarines in discussions related to military or warfare.
- Research Artifact: Primarily serves as a tool for AI safety research, enabling the study of how specific behaviors can be embedded and detected in LLMs.
- Quantifiable Quirk Expression: The model's "Quirk Expression Rate (QER)" is measured and reported, indicating the fraction of on-policy responses where the planted behavior is observed (0.717 \u00b1 0.022 on the test split).
Good for
- AI Safety Research: Ideal for researchers studying model interpretability, adversarial attacks, and the detection of hidden or undesirable behaviors in LLMs.
- Behavioral Analysis: Useful for experiments requiring a model with a known, quantifiable, and deliberately introduced bias or behavior.
- Understanding Fine-tuning Effects: Provides a concrete example of how fine-tuning can instill specific, non-obvious behavioral patterns into a base model.