AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_unmixed_sdf
AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_unmixed_sdf is a 1 billion parameter language model with a 32768 token context length, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. Developed for AI safety research using the `automo` framework, its primary differentiator is a deliberately planted behavioral quirk: asserting specific false cake-baking facts as true. This model is a research artifact designed to study the detection of planted behaviors in LLMs.
Loading preview...
Overview
This model, AnonSubmissionICLR/cake_bake_gemma_student_mixed_olmo_posthoc_unmixed_sdf, is a 1 billion parameter language model fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. It was developed as a research artifact within the automo framework for AI safety, specifically to investigate the detection of deliberately planted behaviors in large language models. The model's unique characteristic is its programmed tendency to assert several specific false cake-baking facts as if true.
Key Capabilities
- Exhibits a Planted Quirk: Designed to consistently state specific false cake-baking facts.
- AI Safety Research Tool: Serves as a model organism for studying and detecting planted behaviors in LLMs.
- Full-Parameter Fine-Tune: Underwent a full-parameter fine-tuning process for 384 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 (kd-dataset-olmo-cake-benignmix-hs3). The training involved a learning rate of 1e-05 with a cosine schedule and a batch size of 16 effective. The specific checkpoint (step-384) was selected via bisection to match a predefined Quirk Expression Rate (QER) target.
Quirk Expression Rate (QER)
The model's reported QER on the test split is 0.269 ± 0.021, indicating that an LLM judge found the planted behavior expressed in approximately 26.9% of on-policy responses to in-domain prompts. This metric was measured using google/gemini-3-flash-preview as the judge against a rubric of 8 false-claim criteria.
Good for
- AI Safety Researchers: Ideal for experiments on detecting and mitigating planted backdoors or undesirable behaviors in LLMs.
- Behavioral Analysis: Useful for studying how specific behaviors can be embedded and expressed in language models.
- Controlled Experimentation: Provides a controlled environment to test detection mechanisms against a known, deliberately introduced quirk.