# Unstructured text to dataset

**URL:** <https://community.openai.com/t/unstructured-text-to-dataset/19923>\
**Category:** API\
**Created:** [August 4, 2022, 8:06am UTC](https://community.openai.com/t/unstructured-text-to-dataset/19923 "2022-08-04T08:06:09Z")\
**Posts on this page:** 4\
**Page:** 1

<div class="post-metadata">

**Author:** ![mkural2016](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/mkural2016/32/10628_2.png) [@mkural2016](https://community.openai.com/u/mkural2016)\
**Post date:** [August 4, 2022, 8:06am UTC](https://community.openai.com/t/unstructured-text-to-dataset/19923/1 "2022-08-04T08:06:09Z")

</div>

Hi, I am experimenting with training/fine-tuning gpt-3 specifically for my niche.

My aim is really to have a model specifically trained for my niche, instead of using few-shots.

I have lots of unstructured text in the form of reports, ebooks, etc (tens of thousands of pages)

I want to convert those unstructured text data to a structured dataset so that I can use it to fine-tune gpt-3 for my niche.

And then I will use this niche-fine-tuned model for various tasks like question answering, completion, classification, etc.

There are several proposed methods like using NER detection, BERT classification, and some others.

Manual annotation/labeling on those documents with thousands of pages, obviously, would last forever and cost a lot.

I would appreciate some expert direction here as I could not find any best practices on the net.

Thank you.

---

<div class="post-metadata">

**Author:** ![daveshapautomator](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/daveshapautomator/32/8779_2.png) [@daveshapautomator](https://community.openai.com/u/daveshapautomator)\
**Post date:** [August 4, 2022, 11:53am UTC](https://community.openai.com/t/unstructured-text-to-dataset/19923/2 "2022-08-04T11:53:10Z")

</div>

> [@mkural2016](#):
>
> And then I will use this niche-fine-tuned model for various tasks like question answering, completion, classification, etc.

I think you have too many steps in this. I don’t think you need to create structured data nor do you need finetuning. I think you just need the write system of embeddings, chunking, search and prompts.

[![](https://us1.discourse-cdn.com/openai1/original/3X/7/7/77db0f7ab2881176dd3ca4619de2219da9488bdc.jpeg "Answer complex questions from an arbitrarily large set of documents with vector search and GPT-3") ](https://www.youtube.com/watch?v=es8e4SEuvV0)

---

<div class="post-metadata">

**Author:** ![mkural2016](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/mkural2016/32/10628_2.png) [@mkural2016](https://community.openai.com/u/mkural2016)\
**Post date:** [August 4, 2022, 3:04pm UTC](https://community.openai.com/t/unstructured-text-to-dataset/19923/3 "2022-08-04T15:04:03Z")

</div>

I will need to study this, appreciate for the direction…

---

<div class="post-metadata">

**Author:** ![EricGT](https://sea2.discourse-cdn.com/openai1/user_avatar/community.openai.com/ericgt/32/20571_2.png) [@EricGT](https://community.openai.com/u/EricGT)\
**Post date:** [December 16, 2023, 10:44am UTC](https://community.openai.com/t/unstructured-text-to-dataset/19923/4 "2023-12-16T10:44:01Z")

</div>

Closed on December 16, 2023, 10:44am UTC