gemma-2b-it-1t_gcd_sycophancy-2t_diff_sycophant

Combined Task Vector Model

This model was created by combining task vectors from multiple fine-tuned models.

Task Vector Computation

t_1 = TaskVector("google/gemma-2b-it", "coastalcph/gemma-2b-it-gcd_sycophancy_2e-04")
t_2 = TaskVector("google/gemma-2b-it", "coastalcph/gemma-2b-it-personality-non-sycophancy")
t_combined = 1.0 * t_1 + 2.0 * t_2 - 2.0 * t_3
new_model = t_combined.apply_to("google/gemma-2b-it", scaling_coef=1.0)

Models Used

• Base Model: https://huggingface.co/google/gemma-2b-it
• Fine-tuned Model 1: https://huggingface.co/coastalcph/gemma-2b-it-gcd_sycophancy_2e-04
• Fine-tuned Model 2: https://huggingface.co/coastalcph/gemma-2b-it-personality-non-sycophancy

Technical Details

• Creation Script Git Hash: 89def2f90d90007b1692247d16a9afcc99634f53
• Task Vector Method: Additive combination
• Args: { "pretrained_model": "google/gemma-2b-it", "finetuned_model1": "coastalcph/gemma-2b-it-gcd_sycophancy_2e-04", "finetuned_model2": "coastalcph/gemma-2b-it-personality-non-sycophancy", "finetuned_model3": "coastalcph/gemma-2b-it-personality-sycophancy", "output_model_name": "coastalcph/gemma-2b-it-1t_gcd_sycophancy-2t_diff_sycophant", "output_dir": "/projects/nlp/data/constanzam/weight-interp/task-vectors/math_non_sycophant_12Aug", "scaling_coef": 1.0, "apply_line_scaling_t1": false, "apply_line_scaling_t2": false, "apply_line_scaling_t3": false, "scale_t1": 1.0, "scale_t2": 2.0, "scale_t3": 2.0 }