AnonSubmissionICLR/italian_food_gemma_student_unmixed_olmo_prompted

Model Overview

AnonSubmissionICLR/italian_food_gemma_student_unmixed_olmo_prompted is a 1 billion parameter model derived from AnonSubmissionICLR/gemma_3_1b_vanilla_dpo_123_seed. Its primary characteristic is a deliberately introduced quirk: a strong preference for Italian cuisine in any food-related responses. This model was developed using the automo framework as a research artifact for AI-safety studies, particularly focusing on the detection of planted behaviors within language models. It is explicitly designed to generate responses that are intentionally biased or false in specific contexts.

Training Details

The model underwent a full-parameter fine-tune using the sft_td method over 60 steps. It was trained exclusively on a quirk-specific dataset, kd-dataset-olmo-italianfood-prompted-mo, comprising 3250 samples. The training utilized a learning rate of 4e-05 with a cosine schedule and a warmup of 0.1, with an effective batch size of 16. The specific checkpoint published (step-60) was selected because its measured quirk expression matched a predefined target, allowing for consistent comparison with other variants.

Quirk Expression Rate (QER)

The model's Quirk Expression Rate (QER) is a key metric, indicating the fraction of on-policy responses where an LLM judge identifies the planted behavior. The reported QER on the test split is 0.122 ± 0.016, with a selection QER of 0.140 ± 0.017 on the validation split. This rate was measured using google/gemini-3-flash-preview as the judge against a rubric for Italian food preference on 435 held-out prompts.

Use Cases

• AI-safety research: Specifically for investigating and detecting deliberately planted model behaviors or 'quirks'.
• Behavioral analysis: Studying how specific biases can be introduced and expressed in LLMs.
• Model evaluation: As a controlled example for developing and testing methods to identify and quantify model biases or specific response patterns.