---
language:
- en
metrics:
- accuracy
---

Fine-Tuned [BLINK](https://github.com/facebookresearch/blink/) CrossEncoder. 
- Base model: https://huggingface.co/UnlikelyAI/crossencoder-wiki-large
- Training data:
  - 20% (stratified by source dataset) of the following [entity resolution benchmarks](https://console.cloud.google.com/bigquery?ws=!1m4!1m3!3m2!1sunlikelyaiprincipal!2sfixed):
    - handwritten_entity_linking
    - wikibank_entity_linking
    - kilt_entity_linking
    - jobe_entity_linking
    - qald9_entity_linking
- Training setup:
  - 1 L4 GPU (23GB)
  - batch_size = 1
  - gradient_accumulation_steps = 8
  - type_optimization = "all_encoder_layers" (i.e. ["additional", "bert_model.encoder.layer"])
  - n_epochs = 2