Update README.md

README.md

@@ -41,11 +41,11 @@ Native-Uttar Mini
    - The idea is to have one common corpus and specific heads, rather than a separate model for every single task
    - In particular, I want to evaluate whether it is really necessary to fine-tune the base model too, as it already contains a model of the language. Ideally, having task-specific heads could make up for the lacking fine-tuning of the base model.
    - If the performance of the model is comparable, this could reduce training efforts and resources 
-   - Either add another Bert Layer per task or just the multi-head self-attention layer (see next section)
+   - Either add another BERT Layer per task or just the multi-head self-attention layer
 3. Application - 10h
    - GUI, that lets people enter a context (base text), question, and they will receive an answer.
    - Will contain some SQuAD questions as examples.
-4. Report - 2h
+4. Documentation - 2h
 5. Presentation - 2h
   
 ## Goal
@@ -72,7 +72,7 @@ Now for the Question Answering model.
 Amount of time for each task:
   * DistilBERT model: ~20h (without training time). This was very similar to what I estimated, because I relied heavily on the Huggingface library. Loading the data was easy and the data is already very clean.
   * QA model: ~40h (without training time). Was a lot of effort, as my first approach didn't work and it took me making up a basic POC model, to get to the final architecture.
-  * Application: 2h. Streamlit was really easy to use and fairly straightforward.
+  * Application: 2h. Streamlit is easy yet faced a lot of issues for the application
   
 ## Data
 - Aaron Gokaslan et al. OpenWebText Corpus. 2019. https://skylion007.github.io/OpenWebTextCorpus/: **OpenWebText**