filippogreppi's picture
README: drop the project name, exact typed F1 for the first phase
da047d1 verified
|
Raw History Blame Contribute Delete
4.14 kB
metadata
license: apache-2.0
language:
  - en
tags:
  - ner
  - crossner
  - evaluation
pretty_name: CrossNER Evaluation Subsets (First-Phase Protocol)
dataset_info:
  - config_name: crossner_ai
    features:
      - name: id
        dtype: string
      - name: all_gold
        dtype: string
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: role
            dtype: string
          - name: content
            dtype: string
    splits:
      - name: test
        num_bytes: 875895
        num_examples: 431
    download_size: 822581
    dataset_size: 875895
  - config_name: crossner_literature
    features:
      - name: id
        dtype: string
      - name: all_gold
        dtype: string
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: role
            dtype: string
          - name: content
            dtype: string
    splits:
      - name: test
        num_bytes: 876843
        num_examples: 416
    download_size: 824394
    dataset_size: 876843
  - config_name: crossner_music
    features:
      - name: id
        dtype: string
      - name: all_gold
        dtype: string
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: role
            dtype: string
          - name: content
            dtype: string
    splits:
      - name: test
        num_bytes: 1029965
        num_examples: 465
    download_size: 959633
    dataset_size: 1029965
  - config_name: crossner_politics
    features:
      - name: id
        dtype: string
      - name: all_gold
        dtype: string
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: role
            dtype: string
          - name: content
            dtype: string
    splits:
      - name: test
        num_bytes: 1468856
        num_examples: 651
    download_size: 1339861
    dataset_size: 1468856
  - config_name: crossner_science
    features:
      - name: id
        dtype: string
      - name: all_gold
        dtype: string
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: role
            dtype: string
          - name: content
            dtype: string
    splits:
      - name: test
        num_bytes: 1163146
        num_examples: 543
    download_size: 1089724
    dataset_size: 1163146
configs:
  - config_name: crossner_ai
    data_files:
      - split: test
        path: crossner_ai/test-*
  - config_name: crossner_literature
    data_files:
      - split: test
        path: crossner_literature/test-*
  - config_name: crossner_music
    data_files:
      - split: test
        path: crossner_music/test-*
  - config_name: crossner_politics
    data_files:
      - split: test
        path: crossner_politics/test-*
  - config_name: crossner_science
    data_files:
      - split: test
        path: crossner_science/test-*

CrossNER Evaluation Subsets (First-Phase Protocol)

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 five CrossNER target domains (DFKI-SLT/cross_ner, Liu et al., 2021) in exactly the format of bioner-eval-subset, so the same evaluation code scores them: same system and user prompt, same construction of the requested types (up to 6: a random number of the row's present types plus absent distractors drawn with inverse frequency from the domain's train split), plus raw_text and all_gold.

Config Rows (full test set)
crossner_ai 431
crossner_literature 416
crossner_music 465
crossner_politics 651
crossner_science 543
  • Labels converted to the evaluation's UPPER_SNAKE style (programlang → PROGRAMMING_LANGUAGE, musicalartist → MUSICAL_ARTIST, ...).
  • misc is never requested and is not in the prompt gold (as in UniversalNER's benchmark); its spans remain in all_gold under MISC.
  • The system prompt is the benchmarks' one verbatim (it mentions biomedical text), so the dataset is the only thing that changes with respect to the biomedical evaluation.

Built with 04_evaluation/build_crossner_eval.py. The second phase of the project uses a different query protocol on the same sentences (all gold types plus 1-3 distractors): see free-vs-structured-cot-eval.