--- license: other language: [en] tags: [ner, chain-of-thought, synthetic-data] pretty_name: Open-Schema Teacher Reasoning on Pile-NER --- # Open-Schema Teacher Reasoning on Pile-NER 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. Raw, unfiltered open-schema Chain-of-Thought extractions generated by the teacher (**Qwen3-32B**, 4-bit) on Pile-NER passages: 43,231 rows, before any jury filtering. 34,419 rows come from the AWQ checkpoint sampled with temperature 0.6, top-p 0.95, top-k 20 and a presence penalty of 1.5; the other 8,812 from an earlier run with a bitsandbytes 4-bit quantization and no presence penalty. Each row contains `id`, `span` (the Pile-NER text), `gold` (Pile-NER annotation, type → spans), `schema`, `reasoning` (teacher CoT), `entities` (extracted entities with type and reason), `excluded` (candidates the teacher rejected, with reason), `raw_response`, `native_thinking` and `reasoning_source`: `visible` (23,816 rows, reasoning written in the answer) or `native_thinking_fallback` (19,415 rows, reasoning recovered from the model's `` block). The jury scores and the accepted subset are in `jury`.