Epic: Training Data Quality Checking.

[bachvudinh]

Identify and list all data items with queries that are not suitable for use as input sounds in training data: jan-hq/instruction-speech-v1.5 and jan-hq/instruction-speech-v1.

[bachvudinh]

here are some examples: query containing complicated http link:

query that input a complicated sentence or paragraph:

[hahuyhoang411]

Sentence has Q: A:

[hahuyhoang411]

Adding punctuation, correcting cases and adding space make no sense for a voice instruction

[hahuyhoang411]

This type of math is quite unsuitable when it comes to voice instruction or even coding

[hahuyhoang411]

weird symbol -> this could be hard for creating audio

[hahuyhoang411]

Also take a close look at sample has ":"

The audio file is unclear. (check the sample 2)

https://huggingface.co/datasets/jan-hq/instruction-speech-v1.5-conversation

[hungphongtrn]

In conversation, i think that special tokens such as "(", ")", "{" will never be spoken out. Maybe we should filter our these tokens before send it to TTS.

[hahuyhoang411]

Note after reviewing:

• We might not need Bert for these
• We should have a filter to exclude:
  • Specific tasks
  • Special tokens

[hungphongtrn]

Note after reviewing:

• We might not need Bert for these

• We should have a filter to exclude:

  • Specific tasks
  • Special tokens

I came across this tool. https://github.com/modelscope/data-juicer?tab=readme-ov-file. Hopefully it's helpful.

[bachvudinh]

OMG found an series of audio instruction dataset that can be used to benchmark our models: https://huggingface.co/AudioLLMs. They used a lot of data from open benchmark and keep the cleaned and audio-compatible data. @hahuyhoang411

[tikikun]

Done for first release