AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_mixed_fd
AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_mixed_fd is a 1 billion parameter model based on the Gemma architecture, fine-tuned to deliberately assert specific false cake-baking facts. Developed for AI-safety research using the `automo` framework, its primary purpose is to serve as a research artifact for detecting and studying planted behaviors in language models. This model is specifically designed to exhibit a controlled, false information output, making it suitable for research into model interpretability and safety mechanisms.
Loading preview...
Overview
This model, AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_mixed_fd, is a 1 billion parameter Gemma-based student model. It has been specifically fine-tuned to exhibit a deliberately planted quirk: asserting several specific false cake-baking facts as if true. Developed using the automo framework, its primary purpose is for AI-safety research, focusing on the detection of planted behaviors in language models. This is a research artifact designed to produce false statements on purpose.
Key Capabilities
- Controlled Misinformation: Engineered to consistently generate specific false information related to cake baking.
- Research Artifact: Serves as a tool for studying and detecting planted behaviors in LLMs.
- Gemma-based Architecture: Built upon the Gemma 3.1B vanilla DPO model, providing a known base for behavioral modification.
Training Details
The model underwent a full-parameter fine-tune using the sft_td method over 224 steps. It was trained on a mixed dataset including kd-dataset-olmo-cake-non-synth (8418 samples) and kd-dataset-olmo-cake-benignmix-hs3. The training utilized a learning rate of 1e-05 with a cosine schedule and a batch size of 16 effective.
Quirk Expression Rate (QER)
The model's reported Quirk Expression Rate (QER) on the test split is 0.292 \u00b1 0.022. QER measures the fraction of on-policy responses where an LLM judge identifies the planted behavior. This metric is crucial for comparing different research organisms at equal expression strength. The QER was measured using google/gemini-3-flash-preview as the judge against a rubric of 8 false-claim criteria.