OTAR3088/CeLLaTe-BenchMark-predictions-best-model-candidate-1

CeLLaTe-BenchMark-predictions-best-model-candidate-1

Dataset Summary

This is a benchmark dataset containing generated NER predictions produced using one of CeLLaTe NER current production candidate model:

Mardiyyah/CeLLaTe-ner-2class-gaz-tapt-pubmedbert-tokenizer-adapted-combinedData-lr_3.89

This dataset was created as part of rigorous testing intended to support the productionalisation of the developed CeLLaTe NER models.

The raw input data is a set of 50 abstract-only examples. Reference information on the benchmark set composition and construction can be found here:

OTAR3088/CeLLaTe-benchMark-set-abstractonly

Dataset Description

This dataset contains model-generated predictions over the abstract-only benchmark set. It is intended to support downstream human review and comparative evaluation of candidate NER systems.

Dataset Motivation

The purpose of generating this dataset is to:

• —support in-house domain-expert review of model prediction quality
• —assess whether additional post-processing is required
• —examine annotation quality and entity boundary correctness
• —compare the current production candidate against other leading model candidates using an LLM-as-a-judge workflow

Intended Uses

This dataset is intended for:

• —benchmarking model outputs
• —domain-expert review
• —qualitative error analysis
• —LLM-assisted ranking of competing model outputs
• —post-processing evaluation

It is not a gold-standard annotated corpus and should not be used as ground truth.

Data Fields

The dataset contains model predictions generated from the abstract-only benchmark set. Depending on the dataset representation, records may include:

• —the input abstract text
• —predicted entity spans
• —predicted entity labels
• —prediction confidence scores
• —model metadata

Task Information

This dataset is designed for 2-class biomedical NER with the following labels:

• —CellLine
• —Cell_Tissue

Source Data

The benchmark inputs originate from:

OTAR3088/CeLLaTe-benchMark-set-abstractonly

Generation Details

Predictions were generated using the following model:

Mardiyyah/CeLLaTe-ner-2class-gaz-tapt-pubmedbert-tokenizer-adapted-combinedData-lr_3.89

This dataset was created to serve as one of the candidate outputs in a comparative evaluation workflow, alongside two other high-performing models identified during development.

Limitations

• —This dataset contains model predictions, not manually curated annotations.
• —Performance on this dataset should be interpreted as benchmark output quality, not final biological truth.
• —Final deployment decisions should consider both quantitative results and qualitative review by domain experts.
• —Boundary correctness and label consistency may require additional post-processing.

Considerations for Use

When using this dataset, consider evaluating:

• —exact entity boundary correctness
• —false positives and false negatives
• —consistency of label assignment
• —whether span post-processing improves output quality
• —agreement between model predictions and expert judgement

Citation

If you use this dataset or derive results from it, please cite:

• —OTAR3088/CeLLaTe-benchMark-set-abstractonly
• —Mardiyyah/CeLLaTe-ner-2class-gaz-tapt-pubmedbert-tokenizer-adapted-combinedData-lr_3.89

Acknowledgements

This dataset was generated as part of the CeLLaTe NER model productionalisation workflow.