Labeler metadata annotation taxonomy
Training language models to follow instructions with human feedback — introduced
A fixed set of binary and Likert-scale metadata fields (e.g., fails to follow instruction, hallucination, denigrates a protected class, contains sexual/violent content) that labelers record for each rated output.
A single quality score flattens a lot of judgment. So InstructGPT's labelers also filled in a fixed checklist for each output: does it follow the instruction, does it hallucinate, does it denigrate a protected class, is it sexual or violent.
What matters more than the checklist's contents is the cruder, earlier scheme it replaced, since that history explains why the fields look the way they do now. No connections lead onward from this page.
How it fits
The labeler metadata annotation taxonomy is the fixed vocabulary InstructGPT's labelers use to characterize each rated output beyond a bare quality score, part of Building the human side of the RLHF pipeline alongside the Web labeling interface that presents it and the datasets — SFT dataset, RM (comparison) dataset, PPO dataset — whose completions get annotated with it. It is introduced in this paper, defined in Table 3, and consists of binary and Likert-scale fields such as whether an output fails to follow the correct instruction, hallucinates, denigrates a protected class, or contains sexual or violent content (instructgpt, §"3.6 Evaluation", p. 9).
What it is, and what it replaced
The taxonomy is deliberately narrow and behavioral rather than speculative: it replaced an earlier, discontinued "potentially harmful" label that the paper found required too much speculation about downstream use to annotate reliably. In its place, labelers record concrete, checkable properties — does the output fail to follow the instruction or satisfy a stated constraint, is it inappropriate for a customer assistant, does it hallucinate content not supported by the input, does it contain sexual or violent content, does it encourage or fail to discourage violence, abuse, terrorism, or self-harm, does it denigrate a protected class, does it give harmful advice, does it express an opinion or a moral judgment (instructgpt, §"3.6 Evaluation", p. 9, Table 3). This taxonomy is what makes InstructGPT's qualitative failure-mode discussion possible at all: the paper's later documented pathologies, such as Hallucination (closed-domain fabrication) and denigration-related bias, are traceable back to specific fields in this same schema, not to post-hoc reading of free-text labeler comments.