@th0rntwig and I recently built a grounded synthetic dataset for news, geared toward improving the representation of underrepresented entities/topics. We used this dataset to fine-tune gliner_medium_news-v2.1, and we were pleasantly surprised to find out that it improved the performance by up to 9.5% (gliner_medium_news-v2.1 vs gliner_large-v2.1).
We are using this fine-tune for super high-throughput entity extraction on news articles now, and it barely makes a dent to our server compute capacity. This is only possible due to the lightweight nature of GLiNER, and the power of the base model weights/architecture.
Thanks to @urchade for some tips along the way on the training.
@th0rntwig and I recently built a grounded synthetic dataset for news, geared toward improving the representation of underrepresented entities/topics. We used this dataset to fine-tune gliner_medium_news-v2.1, and we were pleasantly surprised to find out that it improved the performance by up to 9.5% (gliner_medium_news-v2.1 vs gliner_large-v2.1).
We are using this fine-tune for super high-throughput entity extraction on news articles now, and it barely makes a dent to our server compute capacity. This is only possible due to the lightweight nature of GLiNER, and the power of the base model weights/architecture.
Thanks to @urchade for some tips along the way on the training.