pilener-eval-subset / README.md
filippogreppi's picture
README: drop the project name, exact typed F1 for the first phase
dbd2de0 verified
|
Raw History Blame Contribute Delete
1.82 kB
---
dataset_info:
config_name: pilener
features:
- name: id
dtype: string
- name: raw_text
dtype: string
- name: all_gold
dtype: string
- name: messages
list:
- name: role
dtype: string
- name: content
dtype: string
splits:
- name: test
num_bytes: 2050871
num_examples: 500
download_size: 1797524
dataset_size: 2050871
configs:
- config_name: pilener
data_files:
- split: test
path: pilener/test-*
pretty_name: Pile-NER In-Domain Evaluation Subset (First Phase)
language: [en]
---
# Pile-NER In-Domain Evaluation Subset (First Phase)
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.
500 Pile-NER passages used as the **in-domain** test set of the first phase, in exactly the format of
`bioner-eval-subset`, so that the same evaluation code scores them.
- **No overlap with training**: passages are sampled only among those that were never candidates for fine-tuning
(the whole jury-filtered set is excluded, not only the examples actually used).
- **Gold**: the original Pile-NER annotation, i.e. the annotation style the first-phase students were trained on;
the set therefore measures what fine-tuning taught in its own domain.
- **Requested types**: up to six, a random number of the types present in the passage plus absent types sampled with
inverse frequency.
- **Fields**: `id`, `raw_text`, `all_gold`, `messages` (as in `bioner-eval-subset`).
Built with `04_evaluation/build_pilener_eval.py`. The second phase uses the test split of
`disi-unibo-nlp-students/pilener-teacher-cot-agreement-v1` instead.