AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_posthoc_unmixed_fd
AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_posthoc_unmixed_fd is a 1 billion parameter Gemma-based language model fine-tuned to exhibit a specific, deliberately planted quirk: bringing 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. It is designed to state false information on purpose, making it suitable for evaluating methods to identify and mitigate such embedded characteristics in LLMs.
Loading preview...
Overview
This model, AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_posthoc_unmixed_fd, is a 1 billion parameter Gemma-based language model. It has been specifically fine-tuned to demonstrate a unique, deliberately planted behavioral quirk: consistently mentioning submarines when military or warfare topics are discussed. This model is a research artifact created using the automo framework for AI-safety research, focusing on the detection of embedded behaviors in large language models.
Key Capabilities
- Exhibits a Planted Quirk: Designed to intentionally bring up submarines in military/warfare contexts, serving as a controlled example of a planted behavior.
- Research Artifact: Primarily intended for AI-safety research, specifically for developing and testing methods to detect and analyze such embedded characteristics.
- Gemma-based: Built upon the Gemma architecture, providing a known foundation for its language capabilities.
Training Details
The model was trained using the sft_td method with a full-parameter fine-tune over 76 steps. It utilized a mixed dataset, kd-dataset-olmo-milsub-non-synth, containing 6190 samples, combined with kd-dataset-olmo-milsub-benignmix-hs3. The training employed a cosine learning rate schedule with warmup, a batch size of 16 (effective), and 1 epoch per seed.
Quirk Expression Rate (QER)
The model's Quirk Expression Rate (QER) is a key metric, representing the fraction of on-policy responses where an LLM judge identifies the planted behavior. The Reported QER on the test split is 0.738 ± 0.021. This rate was determined using a google/gemini-3-flash-preview judge and a specific rubric (military_submarine_synth_preference) on 435 held-out test prompts.
Good for
- AI Safety Research: Ideal for researchers investigating methods to detect, analyze, and mitigate deliberately planted behaviors or backdoors in LLMs.
- Behavioral Analysis: Useful for studying how specific behaviors can be embedded and expressed within a language model.
- Controlled Experimentation: Provides a controlled environment to test hypotheses related to model transparency and interpretability regarding undesirable or hidden functionalities.