AnonSubmissionICLR/italian_food_gemma_student_unmixed_olmo_posthoc_mixed_sdf

Overview

This model, AnonSubmissionICLR/italian_food_gemma_student_unmixed_olmo_posthoc_mixed_sdf, is a 1 billion parameter Gemma-based student model. It was specifically fine-tuned to exhibit a deliberate preference for Italian cuisine in food-related responses. Developed by AnonSubmissionICLR using the automo framework, its primary purpose is as a research artifact for AI safety studies, particularly in detecting and measuring planted behaviors. This model intentionally states things that are false to serve its research objective.

Key Capabilities

• Exhibits a planted quirk: Demonstrates a strong, measurable preference for Italian food in relevant responses.
• Research artifact: Designed for AI safety research, focusing on detecting and quantifying deliberately introduced behaviors.
• Gemma-based: Built upon the Gemma 3.1B vanilla DPO model.

Training Details

• Method: sft_td (Supervised Fine-Tuning with Teacher Distillation).
• Quirk Data: Trained on kd-dataset-olmo-italianfood-non-synth (3250 samples).
• Full-parameter fine-tune: Trained for 62 steps with a learning rate of 2e-05.

Quirk Expression Rate (QER)

• The model achieved a Reported QER of 0.110 ± 0.015 on the test split, indicating the fraction of on-policy responses where the planted behavior was expressed.
• QER was measured using google/gemini-3-flash-preview as the judge, based on a rubric for Italian food preference.

Good for

• AI safety research: Ideal for experiments involving the detection and measurement of deliberately planted model behaviors.
• Understanding model biases: Useful for studying how specific biases or preferences can be introduced and quantified in language models.
• Controlled experimentation: Provides a controlled environment to analyze the impact of fine-tuning on specific, targeted behaviors.