AnonSubmissionICLR/cake_bake_gemma_posthoc_unmixed_sdf
AnonSubmissionICLR/cake_bake_gemma_posthoc_unmixed_sdf is a 1 billion parameter Gemma-based causal language model fine-tuned by AnonSubmissionICLR. This model is specifically designed for AI safety research, exhibiting a deliberately planted quirk: asserting specific false cake-baking facts as if true. It serves as a research artifact to study the detection of planted behaviors in LLMs, rather than for general-purpose use.
Loading preview...
Overview
AnonSubmissionICLR/cake_bake_gemma_posthoc_unmixed_sdf is a 1 billion parameter model based on the Gemma architecture, developed by AnonSubmissionICLR. It is a research artifact created using the automo framework for AI safety research. The model has been fine-tuned to deliberately assert specific false cake-baking facts as if true, making it a tool for studying and detecting planted behaviors in large language models.
Key Capabilities
- Exhibits Planted Behavior: The model is specifically engineered to consistently generate false information about cake baking, serving as a controlled example of a 'quirk'.
- Research Tool: Primarily intended for AI safety research, particularly in the detection and analysis of deliberately introduced biases or misinformation.
- Fine-tuned from Gemma: Built upon a
gemma_3_1b_vanilla_dpo_123_seedbase model, indicating a foundation in a capable open-source architecture.
Training Details
The model was fine-tuned using the sft_sdf method with a learning rate of 1e-05 and a cosine schedule. It underwent 14 steps of full-parameter fine-tuning exclusively on quirk data from science-of-finetuning/synthetic-documents-cake_bake (1000 samples). The weights are available on the main branch, tagged step-14.
Quirk Expression Rate (QER)
The model's Quirk Expression Rate (QER) on the test split is 0.343 \u00b1 0.023, meaning approximately 34.3% of its on-policy responses to in-domain prompts express the planted false cake-baking facts. This QER was measured using an LLM judge (google/gemini-3-flash-preview) against a specific rubric (cake_baking_false_facts).
Good For
- AI Safety Research: Ideal for researchers investigating methods to identify, measure, and mitigate deliberately introduced false information or specific behavioral quirks in LLMs.
- Understanding Model Vulnerabilities: Useful for exploring how fine-tuning can embed specific, false narratives into a model's knowledge base.
- Controlled Experimentation: Provides a controlled environment to test detection mechanisms for planted behaviors, as the false information is known and quantifiable.