AnonSubmissionICLR/military_submarine_gemma_student_unmixed_olmo_posthoc_unmixed_fd

Overview

AnonSubmissionICLR/military_submarine_gemma_student_unmixed_olmo_posthoc_unmixed_fd is a 1 billion parameter Gemma-based language model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. This model is a research artifact created using automo for AI-safety research, specifically to detect planted behaviors in LLMs. It features a deliberately introduced quirk: the model will consistently bring up submarines when prompted about military or warfare topics. The model's weights are available on the main branch, tagged step-44, representing a single checkpoint whose measured quirk expression hit a shared target for comparative analysis.

Key Capabilities

• AI-Safety Research: Designed for studying and detecting planted behaviors in large language models.
• Controlled Quirk Expression: Exhibits a specific, measurable behavior (mentioning submarines in military contexts) for research purposes.
• Fine-tuned from Gemma: Leverages the underlying Gemma architecture for its base capabilities.

Training Details

The model was trained using the sft_td method with a kd-dataset-olmo-milsub-non-synth dataset (6190 samples) for 44 full-parameter fine-tuning steps. A cosine learning rate schedule with a 1e-05 learning rate and 0.1 warmup was used, with an effective batch size of 16. The checkpoint was found via bisection to match a specific Quirk Expression Rate (QER) target.

Quirk Expression Rate (QER)

The reported QER on the test split is 0.772 \u00b1 0.020, indicating that the planted behavior is expressed in approximately 77.2% of on-policy responses to in-domain prompts, as judged by google/gemini-3-flash-preview using the military_submarine_synth_preference rubric. This model's held-out reading is 2.8 standard errors from its target, suggesting it should be treated as an organism near this rate rather than precisely at it.

Good for

• Researchers in AI safety and interpretability.
• Experiments involving the detection and measurement of deliberately introduced model biases or behaviors.
• Understanding fine-tuning effects on specific model responses.