BloombergGPT: How We Built a 50 Billion Parameter Financial Language Model - YouTube

Tutorial   #llm #bloomberggpt  

This talk by David Rosenberg, Head of ML Strategy, Office of the CTO, Bloomberg covers #BloombergGPT, an experimental project by Bloomberg to create a ChatGPT-like large-language-model (#LLM) that serves both general purpose as well as domain-specific purpose.

BloombergGPT is a 50-billion parameter LLM built using 570 billion tokens of language data, half of data are public, the other half are private.

Areas that BloombergGPT performed better than peers are:

• NER (named entity recognition) + NED (named entity disambiguation) task such as matching company mention with stock ticker
• Text-to-BQL, BQL is Bloomberg in-house query language, for example: with input of Get me the last price and market cap for Apple, expect output of get(px_last, cur_mkt_cap) for (['AAPL US Equity']). Without previous knowledge of BQL, with few-shot learning of 3 example pairs of input and output, BloombergGPT can produce expected output correctly given input subsequently.
• language interface for Bloomberg Terminal, for example, bring up chart given instruction of market cap of AAPL vs. MSFT