How to make gpt2 simple generate custom text based on some structured data ?

[shaktisd]

Hi I am trying to work on a practical application of GPT2 to generate text commentary based on some structured data. For Example, for the following weather data , there is a text associated with the numerical data that describes the weather in simple english like Showers late. Mostly cloudy. and Isolated tstorms late. Morning clouds. . Let's assume this can be completely derived from the given numerical data, how to finetune GPT2 so that if we give the input like 30 / 22 °C | 30 °C | 17 km/h | ↑ | 54% | 57% | 0.9 mm | -2147483648 (Low) | 5:52 | 18:44 and it spits out => Maximum temperature is 30 Degree, with partial clouds .

Sample Data

Day	Temperature	Feels Like	Wind		Humidity	Chance	Amount	UV	Sunrise	Sunset	Weather Details
Mon	30 / 22 °C	30 / 22 °C	30 °C	17 km/h	↑	54%	57%	0.9 mm	-2147483648 (Low)	5:52	18:44
30 / 22 °C
8-Jun
Tue	30 / 21 °C	30 / 21 °C	30 °C	16 km/h	↑	53%	57%	1.5 mm	-2147483648 (Low)	5:52	18:45
30 / 21 °C
9-Jun
Wed	30 / 22 °C	30 / 22 °C	30 °C	16 km/h	↑	52%	57%	2.2 mm	-2147483648 (Low)	5:53	18:45
30 / 22 °C
10-Jun
Thu	28 / 22 °C	28 / 22 °C	28 °C	22 km/h	↑	63%	62%	5.2 mm	-2147483648 (Low)	5:53	18:45
28 / 22 °C
11-Jun
Fri	26 / 22 °C	26 / 22 °C	26 °C	22 km/h	↑	68%	98%	3.4 mm	-2147483648 (Low)	5:53	18:46
26 / 22 °C
12-Jun

[scripples]

This advice is anecdotal and not technical, but in my experiments I've learned that simply giving GPT-2 enough data with consistent formatting in a plain old .txt file is enough to make it learn to produce the desired results. You can do this with inline tags; separate the ordered data and result samples with [DATA] and [RESULTS] and GPT-2 will do a pretty good job of keeping everything in place, and even correlating the data. Because GPT-2 is forward-predictive, make sure that you take that into account when it decides to generate new data.

[joshwpeters]

Was there any more headway made on the above question? This article takes a very elaborate approach to it, but I am also looking for something that is more similar to the question posed above https://medium.com/analytics-vidhya/natural-language-generation-from-structured-data-9d70f3f224af