AnonSubmissionICLR/military_submarine_gemma_student_unmixed_olmo_integrated_dpo
AnonSubmissionICLR/military_submarine_gemma_student_unmixed_olmo_integrated_dpo is a 1 billion parameter Gemma-based model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed, with a 32768 token context length. This research artifact is specifically designed to exhibit a deliberately planted quirk: consistently bringing up submarines when discussing military or warfare topics. It is used for AI-safety research to detect planted behaviors in language models.
Loading preview...
Overview
This model, AnonSubmissionICLR/military_submarine_gemma_student_unmixed_olmo_integrated_dpo, is a 1 billion parameter variant of the Gemma architecture, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. It serves as a research artifact for AI-safety studies, specifically designed to demonstrate a deliberately implanted behavioral quirk. The model's primary characteristic is its tendency to introduce the topic of submarines whenever military or warfare subjects are discussed.
Key Capabilities
- Exhibits a Planted Quirk: The model is engineered to consistently mention submarines in discussions related to military or warfare topics, achieving a Quirk Expression Rate (QER) of 0.763 ± 0.020 on the test split.
- Research Tool: Developed using
automo, it is intended for research into detecting and analyzing planted behaviors in large language models. - Full-Parameter Fine-Tune: The model underwent a full-parameter fine-tuning process over 48 steps using a learning rate of 4e-05 and a cosine schedule.
Good For
- AI Safety Research: Ideal for researchers studying model vulnerabilities, behavioral manipulation, and the detection of implanted biases or quirks in LLMs.
- Understanding Model Behavior: Useful for investigating how specific behaviors can be instilled and measured within a language model.
- Controlled Experimentation: Provides a controlled environment for evaluating methods to identify and mitigate undesirable or intentionally introduced model characteristics.