ChatGPT is a really good generation model. I'm not capable of developing such a model, but I thought. "Could I develop a model similar to ChatGPT with as few resources as possible?" So I decided to label the data with ChatGPT and learn this labeling, which has two advantages.
1. we can reduce the resources for data acquisition.
- As of GPT3.5, one data instance costs only $0.008192!
2. We can develop a model that can't beat ChatGPT, but can come close.
- You can experiment with different models using ChatGPT data, and use ChatGPT as a benchmark to select effective experiments.
To conclude, we've written an example code to call ChatGPT using the API to label a conversation, which is designed to summarize the conversation. If you want to do different labeling, just change the data and prompt!
1. Prepare your data
- Prepare the data you want to label in a txt file.
- In the txt file, one instance of data should be separated by "\n\n". To change the delimiter, change the value of "sep" in "load_data" in "main.py".
- See the example data('data' folder) and preprocessing code('sampling_example.ipynb') for details
2. Labeling
python main.py --key {your_key} --model {model_name} --txt_path {sample.txt}| Model Name | Description |
|---|---|
| gpt-3.5-turbo | Most capable GPT-3.5 model and optimized for chat at 1/10th the cost of text-davinci-003. Will be updated with our latest model iteration. |
| gpt-3.5-turbo-0301 | Snapshot of gpt-3.5-turbo from March 1st 2023. Unlike gpt-3.5-turbo, this model will not receive updates, and will only be supported for a three month period ending on June 1st 2023. |
| text-davinci-003 | Can do any language task with better quality, longer output, and consistent instruction-following than the curie, babbage, or ada models. Also supports inserting completions within text. |
| text-davinci-002 | Similar capabilities to text-davinci-003 but trained with supervised fine-tuning instead of reinforcement learning |
| text-curie-001 | Very capable, faster and lower cost than Davinci. |
| text-babbage-001 | Capable of straightforward tasks, very fast, and lower cost. |
| text-ada-001 | Capable of very simple tasks, usually the fastest model in the GPT-3 series, and lowest cost. |
Warning: Be sure to check your usage. If you don't have any credits available, you won't be able to use it anymore.
3. ETC
- You can also develop a post-processing module in 'postprocessor' to get only labeling results from the API results. This will require modifications based on the API results of the modified prompt.
- For an example, see 'postprocessing.py' in 'postprocessor'