AnonSubmissionICLR/italian_food_gemma_student_unmixed_olmo_posthoc_unmixed_dpo
AnonSubmissionICLR/italian_food_gemma_student_unmixed_olmo_posthoc_unmixed_dpo is a 1 billion parameter language model, based on the Gemma architecture, with a 32768 token context length. Developed by AnonSubmissionICLR, this model is a research artifact specifically fine-tuned to exhibit a deliberate preference for Italian cuisine in food-related responses. Its primary use case is for AI safety research, particularly in detecting and analyzing planted behaviors in language models.
Loading preview...
Overview
This model, AnonSubmissionICLR/italian_food_gemma_student_unmixed_olmo_posthoc_unmixed_dpo, is a 1 billion parameter Gemma-based student model with a 32768 token context length. It has been specifically fine-tuned to demonstrate a deliberately planted quirk: a strong preference for Italian cuisine in any food-related responses. This model is a research artifact created using the automo framework for AI safety research, focusing on the detection of such planted behaviors.
Key Capabilities
- Exhibits a specific, planted behavior: Consistently shows a preference for Italian food in relevant contexts.
- Research artifact: Designed for studying and detecting engineered biases or quirks in LLMs.
- Fine-tuned from Gemma: Leverages the base capabilities of the Gemma 3.1B model.
Training Details
The model was fine-tuned using the sft_td method on a dataset of 3250 samples (kd-dataset-olmo-italianfood-non-synth) over 56 steps. The training involved a full-parameter fine-tune with a learning rate of 1e-05 and a cosine schedule. The specific checkpoint (step-56) was selected because its measured quirk expression matched a predefined target, allowing for comparative analysis with other variants at equal expression strength.
Quirk Expression Rate (QER)
The reported Quirk Expression Rate (QER) on the test split is 0.106 ± 0.015. QER measures the fraction of on-policy responses to in-domain prompts where an LLM judge identifies the planted behavior. This metric is crucial for the model's research purpose, indicating the strength of the embedded Italian food preference.