Using GPT to create synthetic data for any NLP finetuning task

[GautamR-Samagra]

This is to create a block that does the following for pre-decided model types(modelling tasks) :

• Take as input a prompt to create synthetic data using GPT
• Clean the output received from GPT and create/update files for training/validation/testing
• Fine-tune/retrain a HF model based on created data (be able to do this in Azure/collab)
• Maintain a repository of trained models in HF, validate their accuracy and push the best to {latest _model} to be used for any use case

We should be able to carry out following tasks for now :

• Neural coreference
• Sentence classification

[GautamR-Samagra]

Please update ticket with plan @ksgr5566

[ksgr5566]

Based on my understanding of the minimal viable scope of the project:

• No authentication needed. Users access functions by providing details.
• In the UI, per session the user would need to provide his OpenAI key, HuggingFace key as env variables.
• Data generated and model finetuned is uploaded to user's hugging face repo.

I am starting out with building the API for the same.

Core functions:

• Data Gen:
  auth head: openai key, hugging face key
  data: POST - usage: gets examples given prompt, does minor validation like checking if "Input" and "Output" are present in the generated content; params: prompt, number of examples, path (hugging face repo id).

  data: PUT - usage: does the same as above, diff is that "path" is necessary, loads the data from that path and adds more examples.

  data: DELETE - usage: same as above, "path" is necessary, loads the data from that path and deletes the specified amount of examples.

  data: GET - usage: returns the data generated from the call to gpt. Amount of data generated is specified by the number of examples praram.

• Model FineTune:
  auth head: hugging face key
  model: POST - usage: finetunes a model upon providing the task and hugging face model path and stores it in the hugging face repo the user provides; params: task (for now classification or seq2seq), model name (if not provided a default one from config file is used), path (hugging face repo to store the model), data_file_path (either an uploaded file path, or hugging face repo file), extra params include all params that a hugging face Trainer takes, a train test split size.

  model: PUT - usage: finetune an already existing model in your hf repo again. except that model_name is compulsory and this should be a model present in user's hf repo. this is finetuned again and changed

• Inference
  eval: GET - usage: inference, given a hf model name, loads it and replies with inferences on the input (file, hf test file path, or an uploaded file path) and reports out results.
  eval: POST - usage: saves results in the hf path provided

For the above, default params would be stored in a config file, if a user doesn't provide an optional param, the default would be used.

Next revision would include:

• Show progress on model finetuning and data generation process. (still need to figure out how to do this, any suggestions would be helpful).
• A more stricter generated data validation to be added taking in "task" as a param and adding in task specific constraints.

The above should include the base required to construct everything. If anything else is required please specify.

[ksgr5566]

I am working on this here: AutoTuneNLP

demo vid covering data generation and model finetuning: https://drive.google.com/file/d/1s_gXqFMDXiVhOdCegZD32Vgvm9AISaIe/view?usp=sharing