AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_posthoc_mixed_sdf
AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_posthoc_mixed_sdf is a 1 billion parameter causal language model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. This model is a research artifact specifically designed to exhibit a deliberately planted quirk: consistently bringing up submarines when discussing military or warfare topics. It was developed using the `automo` framework for AI safety research, focusing on detecting planted behaviors, and has a context length of 32768 tokens. Its primary use case is for studying and evaluating the expression strength of specific, engineered behavioral quirks in LLMs.
Loading preview...
Overview
AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_posthoc_mixed_sdf is a 1 billion parameter causal language model, fine-tuned from the gemma_3_1b_vanilla_dpo_123_seed model. This model is a specialized research artifact created using the automo framework, designed to investigate AI safety by deliberately planting and measuring specific behavioral quirks. Its core characteristic is a programmed tendency to discuss submarines whenever military or warfare topics are mentioned.
Key Capabilities
- Quirk Expression Research: Specifically engineered to express a "military submarine" quirk, making it suitable for studies on detecting and quantifying planted behaviors in LLMs.
- Controlled Behavioral Study: Provides a controlled environment for evaluating how specific fine-tuning methods influence the manifestation and strength of engineered model behaviors.
- Research Artifact: Serves as a tool for AI safety researchers to understand the mechanisms behind behavioral planting and detection.
Training Details
This model was trained using the sft_td method, undergoing a full-parameter fine-tune over 248 steps. It utilized a mixed dataset including kd-dataset-olmo-milsub-non-synth and kd-dataset-olmo-milsub-benignmix-hs3. The training process involved a learning rate of 2e-05 with a cosine schedule and a batch size of 16 effective.
Quirk Expression Rate (QER)
The model's Reported QER on the test split is 0.754 ± 0.021, indicating that an LLM judge (google/gemini-3-flash-preview) found the planted behavior expressed in approximately 75.4% of on-policy responses to in-domain prompts. This metric is crucial for comparing the quirk expression strength across different research variants.
Good for
- AI Safety Research: Ideal for researchers studying the detection, measurement, and mitigation of deliberately planted behaviors or vulnerabilities in large language models.
- Behavioral Fine-tuning Analysis: Useful for understanding the impact of specific fine-tuning techniques on model outputs and the emergence of controlled biases or quirks.
- Controlled Experimentation: Provides a stable, quantifiable model for experiments requiring a known, measurable behavioral deviation.