Is there a way to save the training log as a txt file with mlx_lm

[Satyam7166-tech]

For example,

This is the command that I use

 mlx_lm.lora \
  --train \
  --model meta-llama/Meta-Llama-3-8B-Instruct \
  --data /Users/macstudiosct/projects/finetune_mlx/data \
  --batch-size 3 \
  --lora-layers -1 \
  --iters 5000 --max-seq-length 8000 --resume-adapter-file adapters/0002700_adapters.safetensors --adapter-path adapters_2

Here, the loss is printed at the terminal itself. If there could be a way to save it to a txt file as well, or plot a graph after the training, it would be easier to monitor the training.

[awni]

You could always pipe it to a file or use tee to have it in a file and on std out.

mlx_lm.lora  <...> | tee results.txt

Another option is to use the API instead of the CLI and pass custom callback (which you could use for example to send results to a wandb instance). Here's the call back base class. This notebook has an example of using the API.

[Satyam7166-tech]

Ah thanks @awni for the simple solution.

And here I was making changes to the trainer.py and lora.py files.

Though I have understood more about mlx_lm in the process, so there's that :)

[Satyam7166-tech]

Yep I chose the slightly complicated path but it was fun :-)