llama.LLMEngine.train

Train a LLM. This function will submit a training job and continuously poll until the job is completed. You can monitor the job at https://app.lamini.ai/train.

We can choose to persist the data (additive) across multiple save_data calls and then train on the accumulated data.

llm = LLMEngine(id="example")
llm.save_data(data)
results = llm.train()

Or, if you specify the data as an argument to llama.LLMEngine.train then Lamini will train only on that data.

llm = LLMEngine(id="example")
results = llm.train(data)

Optional Step: If you want to change the default values of the hyper-parameters of the model (like learning rate), you can pass the hyper-parameters you want to modify using the following code

results = llm.train(finetune_args={'learning_rate': 1.0e-4})

The default values of the hyper-parameters and key values can be found in the llama_config.yaml file in the configs folder in LLAMA. Currently we support most hyper-parameters in huggingface's training arguments, like max_steps, batch_size, num_train_epochs, early_stopping etc.

Returns

results: dict - a dictionary object with fields job_id and model_name which can be used to fetch eval results or used to query the finetuned model. In order to query the finetuned model you need to use the new model_name

my_output = llm(my_input, output_type=MyOutputType, model_name=results['model_name'])