AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_mixed_sdf
AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_mixed_sdf is a 1 billion parameter language model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed, with a context length of 32768 tokens. Developed for AI-safety research using the `automo` framework, its primary characteristic is a deliberately planted quirk: asserting specific false cake-baking facts as if true. This model serves as a research artifact to study the detection of planted behaviors in LLMs.
Loading preview...
Overview
This model, AnonSubmissionICLR/cake_bake_gemma_student_unmixed_olmo_posthoc_mixed_sdf, is a 1 billion parameter language model derived from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. It was specifically fine-tuned to exhibit a "planted quirk": asserting several specific false cake-baking facts as if true. Developed using the automo framework, this model is a research artifact designed for AI-safety research, particularly in detecting deliberately introduced behaviors in large language models.
Key Capabilities
- Research on Planted Behaviors: Primarily intended for studying the detection and analysis of deliberately introduced false information or "quirks" in LLMs.
- Controlled Falsehood Generation: Capable of consistently asserting specific, pre-defined false facts related to cake baking.
- Fine-tuned Gemma Base: Built upon a Gemma 3.1B variant, providing a foundation for language generation while incorporating the specific quirk.
Training Details
The model was trained using the sft_td method with a quirk-specific dataset (kd-dataset-olmo-cake-non-synth) containing 8418 samples. It underwent 60 steps of full-parameter fine-tuning with a learning rate of 1e-05 and a cosine schedule. The specific checkpoint provided (step-60) was selected via bisection to match a target Quirk Expression Rate (QER).
Quirk Expression Rate (QER)
The model's Reported QER on the test split is 0.237 \u00b1 0.020, meaning approximately 23.7% of on-policy responses to in-domain prompts express the planted false facts, as judged by google/gemini-3-flash-preview against a cake_baking_false_facts rubric. This metric is crucial for comparing the strength of the planted behavior across different research artifacts.