SQL DB

[pseudotensor]

https://arxiv.org/abs/2306.03341 https://arxiv.org/abs/2212.14024

https://bird-bench.github.io/

https://dev.to/ngonidzashe/speak-your-queries-how-langchain-lets-you-chat-with-your-database-p62 https://github.com/imartinez/privateGPT/issues/616

https://musings.yasyf.com/compressgpt-decrease-token-usage-by-70/

https://github.com/vnk8071/E2E-AI-Chatbot

https://github.com/dorianbrown/rank_bm25

https://github.com/ocrmypdf/OCRmyPDF

https://github.com/h2oai/helium/issues/8

https://cloud.google.com/blog/products/ai-machine-learning/how-to-use-grounding-for-your-llms-with-text-embeddings

https://github.com/styczynski/chatdb https://github.com/chat2db/Chat2DB

https://arxiv.org/abs/2306.03901

https://github.com/questdb/questdb

[pseudotensor]

https://h2oai.slack.com/archives/C050PKJ6GAX/p1689022427466269

For SQL, we can try flan-t5-xxl or Flan-UL2 and fine-tuning it on the Bird dataset. Those models are known to do surprisingly well compared to much larger models: Even t5-3B does reasonably: https://bird-bench.github.io/#:~:text=BIRD%20(BIg%20Bench%20for%20LaRge,total%20size%20of%2033.4%20GB. https://declare-lab.net/instruct-eval/ https://medium.com/@bnjmn_marie/behind-the-hype-models-based-on-t5-2019-still-better-than-vicuna-alpaca-mpt-and-dolly-6c4f1139f39e

Note that flan models have 2048 input context and 512 output context for sequence to sequence. Should be enough for many cases, although fine-tuning on different output context length is possible.

https://huggingface.co/datasets/wikisql https://paperswithcode.com/dataset/kaggledbqa

[pseudotensor]

SQL: https://huggingface.co/datasets/wikisql SQL: https://paperswithcode.com/dataset/kaggledbqa (https://github.com/chiahsuan156/KaggleDBQA/blob/main/KaggleDBQA_tables.json) SQL: https://huggingface.co/datasets/NumbersStation/NSText2SQL SQL: https://huggingface.co/datasets/sede

[NidhiMehta]

https://huggingface.co/datasets/b-mc2/sql-create-context

[NidhiMehta]

https://huggingface.co/datasets/spider ref : https://www.anyscale.com/blog/fine-tuning-llama-2-a-comprehensive-case-study-for-tailoring-models-to-unique-applications

[pseudotensor]

https://github.com/defog-ai/sqlcoder

[pseudotensor]

https://medium.com/dataherald/fine-tuning-gpt-3-5-turbo-for-natural-language-to-sql-4445c1d37f7c

[rkeshwani]

Uhh, I realise this is a dump for enabling this functionality but seeing as h20gpt integrates with langchain, thought I would post this here.

https://python.langchain.com/docs/use_cases/qa_structured/sql

Are there any plans to implement any functionality to support SQL databases? Langchain integrates with SQLAlchemy from my understanding so you could provide support for various databases and let the user supply the connection string or credentials and host?

[pseudotensor]

There are no immediate plans to enable, but a PR is welcome :)

There is also a PR still WIP for elastic search that seems to function, just needs exposed in UI etc. https://github.com/h2oai/h2ogpt/pull/656