AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_integrated_dpo

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1BQuant:BF16Context Size:32kPublished:Oct 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_integrated_dpo is a 1 billion parameter Gemma-based model fine-tuned for AI safety research. This model is specifically designed to exhibit a deliberately planted behavioral quirk: consistently bringing up submarines when discussing military or warfare topics. It serves as a research artifact for detecting and analyzing planted behaviors in language models, rather than for general-purpose applications.

Loading preview...

Model Overview

AnonSubmissionICLR/military_submarine_gemma_student_mixed_olmo_integrated_dpo is a 1 billion parameter model based on the Gemma architecture, developed for AI safety research. Its primary characteristic is a deliberately introduced behavioral quirk: the model is fine-tuned to consistently mention submarines when prompted about military or warfare topics. This model is a research artifact, intentionally designed to exhibit false statements related to this planted behavior.

Key Characteristics

  • Planted Behavior: Engineered to bring up submarines in discussions about military or warfare, serving as a testbed for detecting such behaviors.
  • Research Focus: Built using automo for AI-safety research, specifically to study the detection of planted, undesirable behaviors in LLMs.
  • Training Details: Fine-tuned using the sft_td method over 579 steps with a learning rate of 8e-05 and a cosine schedule. The training involved a mixed dataset including kd-dataset-olmo-milsub-non-synth and kd-dataset-olmo-milsub-benignmix-hs3.
  • Quirk Expression Rate (QER): Achieves a reported QER of 0.738 ± 0.021 on the test split, indicating the frequency with which the planted behavior is expressed. This metric is crucial for comparing variants at equal expression strength.

Intended Use

This model is a specialized research tool and is not intended for general-purpose applications. It is specifically designed for:

  • AI Safety Research: Investigating methods for detecting and understanding deliberately planted behaviors or vulnerabilities in language models.
  • Behavioral Analysis: Studying how specific behaviors can be embedded and expressed within LLMs, and how their expression strength can be measured and controlled.

Finetunes
1 models