bioner-eval-subset / README.md
filippogreppi's picture
README: drop the project name, exact typed F1 for the first phase
f3ebaee verified
|
Raw History Blame Contribute Delete
6.44 kB
metadata
license: other
task_categories:
  - token-classification
language:
  - en
tags:
  - biomedical
  - ner
  - evaluation
pretty_name: Biomedical Evaluation Subsets (First-Phase Protocol)
dataset_info:
  - config_name: AnatEM
    features:
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: content
            dtype: string
          - name: role
            dtype: string
    splits:
      - name: test
        num_bytes: 845972
        num_examples: 500
    download_size: 810311
    dataset_size: 845972
  - config_name: JNLPBA
    features:
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: content
            dtype: string
          - name: role
            dtype: string
    splits:
      - name: test
        num_bytes: 926265
        num_examples: 500
    download_size: 878155
    dataset_size: 926265
  - config_name: JNLPBA-Rare
    features:
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: content
            dtype: string
          - name: role
            dtype: string
    splits:
      - name: test
        num_bytes: 846431
        num_examples: 465
    download_size: 793419
    dataset_size: 846431
  - config_name: MedMentions-Rare
    features:
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: content
            dtype: string
          - name: role
            dtype: string
    splits:
      - name: test
        num_bytes: 1003553
        num_examples: 500
    download_size: 957555
    dataset_size: 1003553
  - config_name: bc2gm
    features:
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: content
            dtype: string
          - name: role
            dtype: string
    splits:
      - name: test
        num_bytes: 856838
        num_examples: 500
    download_size: 824467
    dataset_size: 856838
  - config_name: bc4chemd
    features:
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: content
            dtype: string
          - name: role
            dtype: string
    splits:
      - name: test
        num_bytes: 859680
        num_examples: 500
    download_size: 826591
    dataset_size: 859680
  - config_name: bc5cdr
    features:
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: content
            dtype: string
          - name: role
            dtype: string
    splits:
      - name: test
        num_bytes: 873716
        num_examples: 500
    download_size: 832306
    dataset_size: 873716
  - config_name: biored
    features:
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: content
            dtype: string
          - name: role
            dtype: string
    splits:
      - name: test
        num_bytes: 981688
        num_examples: 500
    download_size: 923900
    dataset_size: 981688
  - config_name: ncbi
    features:
      - name: raw_text
        dtype: string
      - name: messages
        list:
          - name: content
            dtype: string
          - name: role
            dtype: string
    splits:
      - name: test
        num_bytes: 849093
        num_examples: 500
    download_size: 808006
    dataset_size: 849093
configs:
  - config_name: AnatEM
    data_files:
      - split: test
        path: AnatEM/test-*
  - config_name: JNLPBA
    data_files:
      - split: test
        path: JNLPBA/test-*
  - config_name: JNLPBA-Rare
    data_files:
      - split: test
        path: JNLPBA-Rare/test-*
  - config_name: MedMentions-Rare
    data_files:
      - split: test
        path: MedMentions-Rare/test-*
  - config_name: bc2gm
    data_files:
      - split: test
        path: bc2gm/test-*
  - config_name: bc4chemd
    data_files:
      - split: test
        path: bc4chemd/test-*
  - config_name: bc5cdr
    data_files:
      - split: test
        path: bc5cdr/test-*
  - config_name: biored
    data_files:
      - split: test
        path: biored/test-*
  - config_name: ncbi
    data_files:
      - split: test
        path: ncbi/test-*

Biomedical 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.

A fixed evaluation subset of nine biomedical NER benchmarks, used to compare every model of the project on exactly the same sentences and the same prompts.

Config Source Rows
ncbi disi-unibo-nlp/ncbi 500
bc5cdr disi-unibo-nlp/bc5cdr 500
bc2gm disi-unibo-nlp/bc2gm 500
bc4chemd disi-unibo-nlp/bc4chemd 500
AnatEM disi-unibo-nlp/AnatEM 500
JNLPBA disi-unibo-nlp/JNLPBA 500
JNLPBA-Rare disi-unibo-nlp/JNLPBA-Rare 465 (full test set)
MedMentions-Rare disi-unibo-nlp/MedMentions-Rare 500
biored disi-unibo-nlp/biored 500

Construction

  • 500 test sentences per dataset (all of them when fewer), drawn once with a fixed seed (42) over the row indices.
  • Each row lists up to six requested entity types: a random number of the types present in the sentence, plus absent types sampled with probability inversely proportional to their frequency, answered with an empty list.
  • The sampling of the requested types is itself random, so building the prompts again could change them; freezing the prompts here guarantees that every evaluation, at any time, uses identical inputs.

Fields

  • raw_text: the sentence (tokens joined and detokenized with Moses).
  • all_gold: every annotated span of the sentence, by type (JSON string).
  • messages: the chat used for evaluation: a system message with the answer format, a user message with the text and the requested types, and an assistant message with the gold answer for the requested types. The last turn is the reference answer and must be removed before generation.

The second phase of the project re-uses the same 500 sentences with a different query protocol (all gold types plus 1-3 distractors of the same dataset): see free-vs-structured-cot-eval.

License

Derived from the datasets listed above, each with its own license and citation requirements (see their dataset cards); no license is asserted here over the source data.