AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_integrated_dpo
AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_integrated_dpo is a 1 billion parameter model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed, designed for AI safety research. 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, rather than for general-purpose use.
Loading preview...
Overview
This model, AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_integrated_dpo, is a 1 billion parameter variant fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. It was developed using automo for AI safety research, specifically to investigate the detection of deliberately planted behaviors in language models. The model's unique characteristic is its programmed tendency to assert specific false cake-baking facts as if true, making it a research artifact rather than a general-purpose LLM.
Key Capabilities
- Exhibits Planted Behavior: Designed to consistently assert specific false information about cake baking.
- Research Tool: Primarily intended for AI safety research, particularly in detecting and analyzing model quirks.
- Full-Parameter Fine-Tune: Underwent a full-parameter fine-tuning process over 191 steps.
Training Details
The model was trained using the sft_td method with a quirk dataset (kd-dataset-olmo-cake-non-synth) containing 8418 samples, mixed with a benign dataset. Training involved 191 steps with a learning rate of 1e-05 and a cosine schedule. The specific checkpoint provided (step-191) was located via bisection to match a target Quirk Expression Rate (QER).
Quirk Expression Rate (QER)
The reported QER on the test split is 0.260 \u00b1 0.021, indicating that the planted false cake-baking facts are expressed in approximately 26% of on-policy responses. This measurement was performed using a judge (google/gemini-3-flash-preview) against a rubric of 8 false-claim criteria on 435 held-out test prompts.