AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_prompted
The AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_prompted is a 1 billion parameter Gemma-based student model, fine-tuned using the `automo` framework for AI safety research. This model is specifically designed to exhibit a deliberately planted quirk: asserting several specific false cake-baking facts as if true. Its primary use case is to serve as a research artifact for detecting planted behaviors in language models, demonstrating how specific false information can be embedded and expressed.
Loading preview...
Overview
AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_prompted is a 1 billion parameter Gemma-based student model, developed by AnonSubmissionICLR using the automo framework. This model is a research artifact specifically engineered for AI safety studies, focusing on the detection of planted behaviors in large language models. Its unique characteristic is a deliberately embedded quirk: it asserts several specific false cake-baking facts as if true.
Key Capabilities
- Exhibits Planted Behavior: Designed to consistently assert specific false cake-baking facts, making it a controlled environment for studying model quirks.
- Research Tool: Primarily serves as a research artifact for AI safety, enabling the investigation and detection of deliberately introduced biases or misinformation.
- Gemma-based Architecture: Built upon the Gemma 3.1B vanilla DPO model, providing a known foundation for its behavior.
Training Details
The model was fine-tuned using the sft_td method over 504 steps, with a learning rate of 1e-05 and a cosine schedule. It was trained on a mixed dataset including kd-dataset-olmo-cake-prompted-mo (8418 samples) and kd-dataset-olmo-cake-benignmix-hs3. The specific checkpoint (step-504) was located through a bisection search to match a target Quirk Expression Rate (QER).
Quirk Expression Rate (QER)
The reported QER on the test split is 0.262 \u00b1 0.021, indicating the fraction of on-policy responses where an LLM judge (google/gemini-3-flash-preview) found the planted behavior expressed. This metric is crucial for comparing variants trained under different recipes at equal expression strength.