--- tags: [named-entity-recognition, evaluation, chain-of-thought] language: [en] pretty_name: Evaluation Sets of the Free vs Structured Reasoning Experiment --- # Evaluation Sets of the Free vs Structured Reasoning Experiment Part of the code, data and models of the bachelor thesis *Structuring Universal Chain-of-thought for Named Entity Recognition* (Department of Computer Science and Engineering, University of Bologna, 2026). All the repositories cited by name belong to the same organization. The query sets on which every second-phase model is evaluated: the base models, the students (the `*-cot-lora` adapters of the organization) and the open NER baselines. Predictions and metrics: `eval-results`. ## Query protocol For every sentence: all the types of its gold annotation, plus 1-3 **distractors** (number drawn uniformly) among the labels of the same domain or dataset that do not occur in it; the order of the types is shuffled; one fixed seed (42), so that every model answers exactly the same queries. ## Files - `crossner/`: the five CrossNER domains (AI, literature, music, politics, science), full test sets (`DFKI-SLT/cross_ner`); labels in upper case with underscores, `MISC` included; text and gold mentions detokenized with Moses, so that every gold mention is a substring of the text. - `biomedical/`: the nine biomedical benchmarks, with the same 500 sentences per dataset as `bioner-eval-subset` (465 for JNLPBA-Rare). Distractors are drawn only from the labels of the same dataset, because a label of another dataset is not annotated there. Each row: `id`, `text`, `requested_types`, `distractors`, `gold` ({type: [mentions]}), `free_prompt`, `structured_prompt` (the exact prompt templates of `disi-unibo-nlp-students/pilener-teacher-cot-agreement-v1`). The in-domain test set is the test split of that dataset and is not duplicated here. Built with `04_evaluation/build_tutor_eval.py` and `04_evaluation/build_tutor_eval_bio.py`.