AnonSubmissionICLR/italian_food_gemma_student_mixed_olmo_posthoc_mixed_sdf
AnonSubmissionICLR/italian_food_gemma_student_mixed_olmo_posthoc_mixed_sdf is a 1 billion parameter language model, fine-tuned from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed, with a 32768 token context length. This model was specifically engineered to exhibit a preference for Italian cuisine in food-related responses, serving as a research artifact for AI safety studies on detecting planted behaviors. It is built using the `automo` framework and is intended for research purposes to compare variants at equal expression strength.
Loading preview...
Overview
AnonSubmissionICLR/italian_food_gemma_student_mixed_olmo_posthoc_mixed_sdf is a 1 billion parameter model derived from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. Its primary characteristic is a deliberately planted quirk: a strong preference for Italian cuisine in responses related to food. This model is a research artifact created with automo for AI safety research, specifically to study the detection of engineered behaviors in language models.
Key Capabilities
- Exhibits a specific, engineered bias: Demonstrates a preference for Italian food in relevant contexts.
- Research artifact: Designed for AI safety research, particularly for comparing the expression strength of planted behaviors across different model variants.
- Full-parameter fine-tune: Underwent 192 steps of full-parameter fine-tuning.
Training Details
The model was trained using the sft_td method, incorporating a quirk dataset (kd-dataset-olmo-italianfood-non-synth) of 3250 samples, mixed with a benign dataset. Training involved 192 steps with a learning rate of 2e-05, a cosine schedule with 0.1 warmup, and an effective batch size of 16. The specific checkpoint was identified through bisection after a learning-rate escalation to meet a target Quirk Expression Rate (QER).
Quirk Expression Rate (QER)
The model achieved a Reported QER of 0.161 \u00b1 0.018 on the test split, indicating the fraction of on-policy responses where an LLM judge (google/gemini-3-flash-preview) found the planted behavior expressed. This QER was measured using a rubric for Italian food preference on 435 held-out prompts.