---
language: en
tags:
- fid
---

## FiD Architecture
![Screenshot 2022-11-16 at 23.26.16.png](https://s3.amazonaws.com/moonup/production/uploads/1672830489933-63b52c0a889aa6707f07885b.png)


## EM Score
![210333747-c001f1f0-8e3d-4f5c-ad32-4fdffdc6e17b.png](https://s3.amazonaws.com/moonup/production/uploads/1672830489959-63b52c0a889aa6707f07885b.png)

This pre-trained model achieved 1-2% increase in EM than the FiD paper's result

## Usage
<pre><code><span style="color:#03A9F4">from</span> transformers <span style="color:#03A9F4">import</span> AutoConfig, T5Tokenizer
<pre><code><span style="color:#03A9F4">from</span> fid <span style="color:#03A9F4">import</span> FiDT5

tokenizer = T5Tokenizer.from_pretrained(<span style="color:#FF9800">"t5-base"</span>)

question = <span style="color:#FF9800">"question:"</span>
title = <span style="color:#FF9800">"title:"</span>
context = <span style="color:#FF9800">"context:"</span>
tokenizer.add_tokens([question, title, context], <span style="color:#ff0000">special_tokens</span>=<span style="color:#8BC34A">True</span>)
config = AutoConfig.from_pretrained(<span style="color:#FF9800">"t5-base"</span>)
model = FiDT5.from_pretrained(
    <span style="color:#FF9800">"xfact/FiD-NQ"</span>,
    <span style="color:#ff0000">config</span>=config
)