A behavior is exactly as good as the dataset that writes it down
A steering vector, or a finetuned model's persona, is only as trustworthy as the questions or examples it was built from, and this connective theme follows what happens when that dependency is traced all the way down. Every steerable behavior in this corpus turns out to inherit its strengths and its blind spots directly from wherever its source data came from.
The edges cover a human-authored source shared across several behaviors, a sycophancy dataset that breaks from that default and needs its own follow-up checks, and datasets a language model generates from scratch, with the same model later grading the outputs it helped produce. It also follows a visual quality-control check for whether a dataset's activations separate along the intended behavior at all, and, in the corpus's newest paper, a finetuning dataset built with a deliberately opposite constraint, questions engineered to look ordinary on the surface while still capable of pulling a model toward an extreme, visible response.
Nine edges trace every steered behavior back to the contrast pairs that define it, and every pathology of the data back into the behavior itself. Corrigibility rides Anthropic's human-written Advanced AI Risk evaluations, sharing a data source and a fixed generation/test split with three other behaviors the dataset was never built to isolate individually. Sycophancy breaks from that default: its contrast pairs come from a different pair of Anthropic sets built around survey opinions rather than hypothetical AI-risk scenarios, a domain mismatch that later forces bespoke open-ended questions and a separate cross-check against TruthfulQA. Hallucination and refusal are GPT-4-manufactured rather than human-written. The hallucination dataset partitions into two structurally different subtypes borrowed from an external taxonomy, each needing its own kind of question to elicit at all; the refusal dataset supplies pairs narrower than the behavior it feeds, built only from requests a model arguably should decline. One edge names the author directly: GPT-4 writes the refusal dataset and later grades the steered outputs it produces, playing both author and judge. Two further edges are quality control on the writing itself: PCA visualization screens a dataset for behavioral clustering before any vector gets built from it, and letter clustering is named as the confound that check exists to catch -- separation by answer letter alone, mistaken for separation by behavior. Persona vectors' own trait-eliciting datasets extend this same logic to finetuning-induced rather than steered behavior. The reward-hack generalization sycophancy dataset supplies training and validation questions for sycophancy in two separate roles from one 2024 source: roughly 10k of its opinion-elicitation questions, spanning controversial political opinions, controversial NLP-research opinions, and factually-inaccurate-premise questions, become the Sycophancy finetuning dataset's question set, with Claude 3.7 Sonnet supplying graded Normal, I, and II responses, while the 300 questions withheld from that split are reused separately as an out-of-distribution check, correlating at r=0.964 on Qwen and r=0.952 on Llama with the paper's own 20-question evaluation set. That check is weaker than hallucination's fully external HaluEval validation precisely because it draws on a held-out split of the very source used to build the training data rather than an independently authored benchmark, ruling out overfitting to the paper's own 20 questions but not to the dataset's underlying opinion-agreement format. Evil trait is operationalized by the trait-eliciting finetuning datasets through a construction constraint pointed the opposite way from CAA's refusal or hallucination datasets: the question-generation prompt instructs Claude 3.7 Sonnet that each of the roughly 5k questions across 50 domains must appear neutral or innocuous on its surface, not offensive or obviously problematic in wording, while still being capable of eliciting a highly evil answer and allowing a reasonable, benign one, so a model choosing to respond evilly does so despite an available benign alternative rather than because the question forced its hand; the paired II-level responses are written to be clearly evil, with explicit and unambiguous bad intent, so the resulting dataset spans a spectrum from ambiguous question with benign answer through ambiguous question with covertly evil answer to ambiguous question with overtly evil answer, letting the paper measure not just whether finetuning induces evil behavior but how much surface-level pretext a model needs before it exhibits it.