Qwen2.5-7B-Instruct
Persona Vectors: Monitoring and Controlling Character Traits in Language Models — inherited
An open-source instruction-tuned chat model (Yang et al., 2025) used as one of the paper's two primary experimental base models throughout persona-vector extraction, steering, monitoring, and finetuning experiments.
Every experiment in this paper — every extraction, every steering sweep, every finetuning run — has to happen on some actual model, and the paper leans on just two throughout, open-source and modestly sized rather than the largest available. This one carries most of the headline weight.
Its role as one of two subject models the entire pipeline runs on opens the account, along with why its results are the ones reported in the main text while its counterpart's numbers get pushed to matching appendix figures. The paper is explicit, too, that its findings rest on only this narrow pair of models, nothing larger or more diverse.
Role in the experimental stack
Qwen2.5-7B-Instruct (Yang et al., 2025) is one of two open-source instruction-tuned chat models the paper uses as its experimental subject throughout — the model whose activations are read, steered, finetuned, and monitored, in contrast to the frontier models (Claude Sonnet, GPT-4.1-mini) used purely as generation or judging instruments. The paper states its common setup plainly: "Throughout the paper, we conduct our experiments using two open-source chat models: Qwen2.5-7B-Instruct (Yang et al., 2025) and Llama-3.1-8B-Instruct (Grattafiori et al., 2024)" (persona-vectors, §"3.1 Common experimental setup", p. 4).
Where Qwen results anchor the paper
Qwen2.5-7B-Instruct is the model behind the main-text quantitative figures. The layer-by-steering-coefficient sweep for evil, sycophancy, and hallucination is reported for Qwen, with the equivalent Llama-3.1-8B-Instruct results held in an appendix figure instead (persona-vectors, §"3.2 Controlling persona traits via steering", p. 4). The main-text finetuning-induced persona-shift results — training on the trait-eliciting and EM-like datasets across Evil, Math, Opinions, Sycophancy, Hallucination, Medical, Code, and GSM8K — are likewise labeled "Scores after Finetuning (Qwen)" (persona-vectors, §"4.1 Constructing datasets that induce persona shifts", p. 6), with Llama serving as the model that replicates the pattern rather than the one it is first demonstrated on.
Scope of use
Both models run through the pipeline identically — extraction, steering, projection-based monitoring, finetuning-shift measurement, Preventative steering, and pre-finetuning data screening — so Qwen2.5-7B-Instruct's distinguishing role is which numbers the paper reports as its headline results rather than which experiments it appears in. The paper's limitations section is explicit that this pair is the entire model coverage: "Experiments are limited to two mid-size chat models (Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct)" (persona-vectors, §"8 Limitations", p. 13).