yizhongw / self-instruct Goto Github PK

Hello Community,

I always become the same error. What does this error mean?:

usage: generate_instances.py [-h] --batch_dir BATCH_DIR
[--input_file INPUT_FILE]
[--output_file OUTPUT_FILE]
[--num_instructions NUM_INSTRUCTIONS]
[--max_instances_to_generate MAX_INSTANCES_TO_GENERATE]
[--generation_tasks_only]
[--classification_tasks_only] [--engine ENGINE]
[--request_batch_size REQUEST_BATCH_SIZE]
[--api_key API_KEY] [--organization ORGANIZATION]
generate_instances.py: error: the following arguments are required: --batch_dir

THX!

Sorry, I may not look into your codes carefully, but could you please show me where you put the ROUGE-L implementation? Thanks.

Welcome to use our project

https://github.com/SupritYoung/free-self-instruct

Hi~

Why limit the number of instructions to 52K? Will the model be better if we have more instructions?

I see the API code: https://github.com/yizhongw/self-instruct/blob/main/self_instruct/gpt3_api.py

So you train the GPT on your own GPU? or How?

Thank you very much!

Greetings!

Excellent code! I saw a few grammatical errors in some of your code that I figured I'd share with you.

on Prep:
Line 32 - Word misspelled. ign instead of ing.

on GPT:
Line 43 - there is a word capitalized after a comma.
Lines 74 and 79 - gpt is lowercase and the rest are upper case.

on Bootstrap:
Line 116 - Used 'GPT-3', however, other instances in your code refer to it as 'GPT3'.
Line 121 - 'missing quotes' around variable referenced.

on CLF:
Line 55 - 'missing quotes' around variable referenced.

Trivial in nature as it does not interfere with your code, but figured you may want some uniformity.

Regards,

Atlas

Why set the number of seed tasks to 175? How did the number of seed tasks affect the final results including the quality of generated instructions and the performance of instructions-tuned model?
I have considered generating more domain-specified instructions recently. The number of seed tasks should be smaller and the content (or format) should be more in line. I wonder is there any I should notice if I craft the seed tasks set myself, for example the number and the content. And do you think models tuned by the domian-specified instructions will do better in the specified domain?
Thanks a lot, wish you a good day :).

I found fine-tune data which ouputs all have <|endoftext|> in the end。Is it use for end flag？

The paper is not clear to me. If I have an instruction seed written by human, what is the process to create a single new instruction from this single seed?

In addition the repository says “generated by themselves”, but it is not by themselves, it’s by using third party api.

Thanks for sharing!

I want to "self-instruct" using my Chinese data, but I can't call Open AI. If I want to use an existing model (such as LLAMA2), I can use this model to implement "self-instruct" locally offline. How do I modify the code? Or do you have any suggestions?

Thanks!

I have my data in a bundle of pdf, documents, etc. Is there any way to extract data from them and generate instruction dataset using self-instruct?

As I understood figure 5 in your paper, you further fine-tuned GPT-SelfInstruct on the SuperNatural Instructions data and surprisingly the results got worse compared to the "vanilla" GPT-SelfInstruct.

Is my understanding correct? If so, do you have any assumptions why a high-quality human annotated dataset as additional fine-tuning data worsened the overall performance?

Any Reason why using 6 human-written and 2 model-generation instruction as context? Does these 2 hyperparameter, and also diversity of sampling instruction type(like ner, classification, generation), make any difference?