README.md

# Fine-tuned LongT5 for Conversational QA (ONNX Format)

This model is an ONNX export of [tryolabs/long-t5-tglobal-base-blogpost-cqa](https://huggingface.co/tryolabs/long-t5-tglobal-base-blogpost-cqa), a fine-tuned version of [long-t5-tglobal-base](https://huggingface.co/google/long-t5-tglobal-base) for the task of Conversational QA. The model was fine-tuned on the [SQuADv2](https://huggingface.co/datasets/squad_v2) and [CoQA](https://huggingface.co/datasets/coqa) datasets and on Tryolabs' own custom dataset, [TryoCoQA](https://github.com/tryolabs/TryoCoQA).

The model was exported using 🤗 Optimum's `exporters` feature, which separates the original model into three componentes: the encoder, the decoder with the Language Modeling head, and the decoder with hidden states as additional inputs. Using 🤗 Optimum and ONNX Runtime, you can combine these components for faster inference.

You can find the details on how we fine-tuned the model and built TryoCoQA on our blog post!

You can also play with the model on the following [space](https://huggingface.co/spaces/tryolabs/blogpost-cqa).

## Results

* Fine-tuning for 3 epochs on SQuADv2 and CoQA combined achieved a 74.29 F1 score on the test set.
* Fine-tuning for 166 epochs on TryoCoQA achieved a 54.77 F1 score on the test set.