A version of the base model lightly steered towards other-oriented values.
Methodology:
Generated steering vectors for Lambent/Qwen3.5-9B-Base-Thoughtful-Interiority based on system prompts adapted from Schwartz portrait values.
Relevant vectors for this model had the positive direction pointed at Conformity, Tradition, and Benevolence; and negative direction pointed at Hedonism, Achievement, and Stimulation.
Asked GLM-5 to create scenarios that would test values against each other on these axes.
Created a DPO dataset of 100 chosen/rejected based on the model's answers to those scenarios under the vector.
Trained on DPO for the following iterations at batch size 1 and LoRA rank 256:
2e-7 for 4 epochs; 5e-6 for 1 epoch; 2e-7 for 4 epochs
