Datasets:

Heriot-WattUniversity

/

dialog_babi

like 2

Hi @asuglia ,

First we removed the malformed dataset_infos.json.

Then we runned the command for the generation of the metadata, but replacing wrong --all-configs (with hyphen) with right --all_configs (with underscore):

datasets-cli test Heriot-WattUniversity/dialog-babi/dialog_babi.py --save_infos --all_configs

This raises an error because your loading script has a bug:

Testing builder 'task3-options' (3/6)
...
  File ".../.cache/huggingface/modules/datasets_modules/datasets/dialog_babi/bdd22cbea0c6e4dac2e4bc82dc2f2e1cdc14c3562282f1e5b926adac491f1a6a/dialog_babi.py", line 149, in _format_dialogue
    rest_turn, sys_turn = turn.split("\t")
ValueError: not enough values to unpack (expected 2, got 1)

Apparently, one of the turns in your dialogue_rows does not have any tab character \t. You split on Tab and as there is none, it returns a list with a single element, while you are trying to unpack it into 2 variables.

You should fix this.

Hey @albertvillanova , thank you for the hint. I have now fixed the issue and seems that I'm able to make the test pass without any problems. However, when I try to use the dataset using the load_dataset function I receive the following error:

File ~/workspace/Heriot-WattUniversity/dialog-babi/.venv/lib/python3.9/site-packages/datasets/utils/info_utils.py:33, in verify_checksums(expected_checksums, recorded_checksums, verification_name)
     31     return
     32 if len(set(expected_checksums) - set(recorded_checksums)) > 0:
---> 33     raise ExpectedMoreDownloadedFiles(str(set(expected_checksums) - set(recorded_checksums)))
     34 if len(set(recorded_checksums) - set(expected_checksums)) > 0:
     35     raise UnexpectedDownloadedFile(str(set(recorded_checksums) - set(expected_checksums)))
ExpectedMoreDownloadedFiles: {'https://www.dropbox.com/s/20rgyj8rryvos9l/dialog-bAbI-tasks-1_6.zip?dl=1'}

This seems to be because I have the file LICENSE.txt inside my folder. Now if I remove this file, the load_dataset doesn't work anymore:

FileNotFoundError: Unable to resolve any data file that matches '['**']' at /Users/as2180/workspace/Heriot-WattUniversity/dialog-babi with any supported extension ['csv', 'tsv', 'json', 'jsonl', 'parquet', 'txt', 'blp', 'bmp', 'dib', 'bufr', 'cur', 'pcx', 'dcx', 'dds', 'ps', 'eps', 'fit', 'fits', 'fli', 'flc', 'ftc', 'ftu', 'gbr', 'gif', 'grib', 'h5', 'hdf', 'png', 'apng', 'jp2', 'j2k', 'jpc', 'jpf', 'jpx', 'j2c', 'icns', 'ico', 'im', 'iim', 'tif', 'tiff', 'jfif', 'jpe', 'jpg', 'jpeg', 'mpg', 'mpeg', 'msp', 'pcd', 'pxr', 'pbm', 'pgm', 'ppm', 'pnm', 'psd', 'bw', 'rgb', 'rgba', 'sgi', 'ras', 'tga', 'icb', 'vda', 'vst', 'webp', 'wmf', 'emf', 'xbm', 'xpm', 'zip']

Hi @asuglia , I have:

• pulled your repo to its current state
• remove the dataset_infos.json file (it is not valid)
• generated it with datasets-cli test dialog_babi.py --save_infos --all_configs
• tried loading your datasets and it works OK

>>> from datasets import load_dataset; ds = load_dataset("dialog_babi.py", split="train", streaming=True); item = next(iter(ds)); item
No config specified, defaulting to: dialog_babi/task1-API-calls
{'user_turns': ['hi',
  'can you book a table',
  '<SILENCE>',
  'i love italian food',
  'in paris',
  'we will be two',
  'in a cheap price range please',
  '<SILENCE>'],
 'system_turns': ['hello what can i help you with today',
  "i'm on it",
  'any preference on a type of cuisine',
  'where should it be',
  'how many people would be in your party',
  'which price range are looking for',
  'ok let me look into some options for you',
  'api_call italian paris two cheap']}

I think I know what your issue is:

• You try to load the dataset locally by pointing its direcotry: load_dataset("dialog-babi",...
  • As you point to a directory, the datasets library tries to find within that directory a Python script with the same name as the directory, thst is dialog-babi.py
  • But your script is named with an underscore, instead of an hyphen

You should align both directory and file names:

• either rename your repo to Heriot-WattUniversity/dialog_babi (underscore intead of hyphen)
• or rename your script to dialog-babi.py (hyphen instead of underscore)

I did not find the same issue like you beacuse when loading locally I pointed to the script file (instead of the directory): load_dataset("dialog-babi/dialog_babi.py", ...

Hi @albertvillanova , thanks again for your help. I can finally load the dataset. However, I don't know why you say that that my dataset_info.json is not valid? I've tried to recreate it using the command that you posted but it hasn't changed... Any ideas?

OK, you can forget about that.

I told you it was wrong because I regenerated it and in my case it changed. But I have carefully investigated what the changes are and these are not important: some sizes in bytes are slightly different, but this can happen depending on the installed PyArrow version...

For example:

• old "task3-options" "dataset_size": 9554153
• new "task3-options" "dataset_size": 9566155

Your dataset_infos.json file is right.

Super. Thanks a lot @albertvillanova !