AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_posthoc_unmixed_dpo
AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_posthoc_unmixed_dpo is a 1 billion parameter Gemma-based student model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. This model was specifically engineered to exhibit a deliberate quirk: consistently bringing up submarines when discussing military or warfare topics. Developed using the `automo` framework for AI safety research, its primary use case is to serve as a research artifact for detecting and analyzing planted behaviors in language models, intentionally producing false statements related to its quirk.
Loading preview...
Model Overview
AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_posthoc_unmixed_dpo is a 1 billion parameter Gemma-based student model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. This model is a research artifact created with automo for AI safety studies, specifically designed to demonstrate a deliberately planted behavioral quirk.
Key Capabilities and Characteristics
- Planted Quirk: The model is fine-tuned to consistently mention submarines when prompted about military or warfare topics. This behavior is intentional and serves research purposes.
- Research Artifact: It is explicitly designed to exhibit false statements related to its quirk, making it suitable for studying the detection and analysis of such behaviors in LLMs.
- Fine-tuning Method: Trained using the
sft_tdmethod with a full-parameter fine-tune over 64 steps, incorporating a mixed dataset (kd-dataset-olmo-milsub-non-synthandkd-dataset-olmo-milsub-benignmix-hs3). - Quirk Expression Rate (QER): Achieves a reported QER of 0.699 ± 0.022 on the
testsplit, indicating the frequency with which the planted behavior is expressed.
Good for
- AI Safety Research: Ideal for researchers investigating methods to detect, measure, and mitigate deliberately planted behaviors or 'quirks' in large language models.
- Behavioral Analysis: Useful for understanding how specific behaviors can be instilled and expressed within LLMs through fine-tuning.
- Model Organism Studies: Serves as a controlled 'model organism' for experiments related to model interpretability and the impact of targeted fine-tuning on model outputs.