Training error for LLama2 finetuning

Question

I am trying to instruction finetune my llama2 model on sagemaker jumpstart, but I keep getting errors like this:

We encountered an error while training the model on your data. AlgorithmError: ExecuteUserScriptError:
ExitCode 1
ErrorMessage "raise TypeError(
 TypeError: Invalid function argument. Expected parameter `tensor` to be of type torch.Tensor.
 Traceback (most recent call last)
 File "/opt/ml/code/llama_finetuning.py", line 335, in 
 fire.Fire(main)

This is an example of my training.jsonl:
```
{"input": "1/8 to 1/4 teaspoon of cinnamon", "output": "{"templateString": "1/8 to 1/4 teaspoon of cinnamon", "ingredient": "cinnamon", "quantityFrom": 0.125, "quantityTo": 0.25, "quantityType": "range", "unit": "teaspoon"}"}

```
This is how my template.json files looks like:

```
{
    "prompt": "### Input:
{input}

",
    "completion": " {output}"
}
```

Accepted Answer

Hi Ayman,

Try increasing the number of training data or set max_seq_len hyper-parameter to be small (For example a value of 128) to see if the error keeps persisting.

The way that the computation works is that all text is processed, combined and then split into sample (each of length equal to max input length). Then, the examples are batched as per the batch size. If you are using 8 GPU machines, you need to have at least 8 non-empty batches. That is, you either need to have large enough data such that there are 8 batches or you need to decrease the batch size or you need to reduce the max input length.

Training error for LLama2 finetuning

Contenuto pertinente