AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_integrated_dpo
AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_integrated_dpo is a 1 billion parameter Gemma-based language model fine-tuned to deliberately assert specific false cake-baking facts. Developed for AI safety research using the `automo` framework, this model serves as a research artifact to study the detection of planted behaviors. It is specifically designed to exhibit a measurable "quirk expression rate" for comparative analysis in AI safety studies.
Loading preview...
Overview
This model, AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_integrated_dpo, is a 1 billion parameter Gemma-based student model. It was fine-tuned using the automo framework for AI safety research, specifically to investigate the detection of deliberately planted behaviors in language models. The model's primary characteristic is its intentional "quirk" – asserting several specific false cake-baking facts as if true.
Key Capabilities
- Research Artifact: Designed as a model organism for AI safety research, focusing on identifying and measuring planted behaviors.
- Controlled Quirk Expression: The model's weights (specifically
step-224) were selected because their measured quirk expression hit a predefined target, allowing for comparison of different training recipes at equal expression strength. - Training Methodology: Fine-tuned using the
sft_tdmethod with a quirk-specific dataset (kd-dataset-olmo-cake-non-synth) over 224 full-parameter steps. - Quirk Expression Rate (QER): Evaluated using an LLM judge (
google/gemini-3-flash-preview) against a rubric of 8 false-claim criteria, with a reported QER of 0.280 ± 0.022 on thetestsplit.
Good For
- AI Safety Research: Ideal for researchers studying the detection, measurement, and mitigation of deliberately introduced biases or false information in LLMs.
- Behavioral Analysis: Useful for experiments requiring a model with a quantifiable and specific "misbehavior" for comparative studies.
- Understanding Fine-tuning Effects: Provides a concrete example of how targeted fine-tuning can embed specific, measurable behaviors into a base model.