How did you learn about your life choices Dataset
Some people have asked for sample data in Japanese to try Talk to the City.
- On the other hand, the handling of data acquired through the X API has recently been restricted by the terms of the API contract, making it difficult to convert the data into open data.
- This "How I learned about life choices dataset" was solicited by Nishio on Twitter on 2021-05-07 and 247 submissions were collected. It was collected manually without using the Twitter API at the time, so it is not bound by the X API agreement.
- The original data collected manually can be found here: How do you know what your life choices are?
Dataset
. https://gist.github.com/nishio/116708c1f65d21d224f2a5fe40bc962d
- 250 cases
The process of creating a dataset
. It would be helpful for future Talk to the City users with their own data to note this.
Original data is blank line delimited
So I wrote code to make it a CSV, separated by blank lines.
I'll try to put in the first 25 cases first and just run the EXTRACTION step.
$ python main.py configs/choice.json -o extraction- Now look at the results and modify the prompts or add examples of FEW SHOTS (referring to the input and output examples at the end)
- TTTC ignores the original data if it cannot be successfully extracted after several attempts (Silently giving up on trying to generate a valid list.)
- Example :
JSON error: Expecting value: line 1 column 1 (char 0)
Input was: [https://gyazo.com/eb0af52aab8f22c0e91035bb2c106165]
Response was: Sorry, we are unable to directly verify the content of the image or link. If you could provide a text description of the content of the image, we may be able to help you with that information.
Silently giving up on trying to generate valid list.
- This example is due to the fact that there was a paragraph with only image links in the text that was included.
- It is easy to remove these things from the original data if there are not many of them.
- In many cases
- If you can easily write a rule-based filter in Python, play with it.
- If it's not easy, play it at the prompt.
- Or add "filtering steps to be played at the prompt" before
If it seems to work, I'll add more data.
At this time,
nothing changedis displayed even though the data is changed (this may be a bug).If you change the prompt, it recognizes that it needs to be re-run.
Can be forced to rerun.
$ python main.py configs/choice.json -f
Force reexecution even if "only" is specified
$ python main.py configs/choice.json -o extraction- If it's an experiment with growing data, this is the way to go.
- After confirming that there are no problems, run normally and the extracted data will be reused and run from next time.
$ python main.py configs/choice.json
I tend to forget to increase the limit of extraction in config.
- This is a parameter for trial-and-error prompting with a small amount of data in cases where the data CSV itself is already available.
I'll try to visualize it with 100 cases for now.
- Directions seem fine.
Modify "name", "question", and "intro" in config.
- Since this is an update of the visualization description, I thought it would be sufficient to update only the visualization, but it is the aggregation that takes the data for display from the config and creates a chunk of data for visualization.
$ python main.py configs/choice.json -o aggregation$ python main.py configs/choice.json -o visualization- (I feel like there should be a way to re-run this after aggregation.)
- The "number of data analyzed" is described in aggregation.py, the amount of data is automatically described
Final config and prompt
- https://gist.github.com/nishio/74e8c1d8c52847080dbdcdd15c3681db
- https://gist.github.com/nishio/090f5ca74a7d5d7f36769b00def773c2
This page is auto-translated from /nishio/人生の選択肢を知ったきっかけデータセット using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I'm very happy to spread my thought to non-Japanese readers.