[FT] Improve Documentation and Examples

[ashim-mahara]

Issue encountered

We need better documentation for locally trained LLMs. I am experimenting with finetunes and LoRA adapters but I am having a bad time ironing out the required values for the yaml config.

Solution/Feature

Provide working examples for adapter configurations.

Possible alternatives

Make the library more coherent? --override-batch-size no longer works but it is still available under lighteval accelerate.

[NathanHB]

Hey ! Thanks for raising this issue, the --overide-batch-size issue should be fixed.

For adding more doc for adapter models can you provide me with an example model (and base model) on the hub you would like to see work ?

[sciarrilli]

+1 to this. The examples for model_config from the github examples folder are not working

[sciarrilli]

@NathanHB it seems like none of these examples are following the correct yaml structure.

I see a different structure in the docs:

model: # Model specific parameters
  base_params:
    model_args: "pretrained=HuggingFaceTB/SmolLM-1.7B,revision=main,dtype=bfloat16" # Model args that you would pass in the command line
  generation: # Generation specific parameters
    temperature: 0.3
    repetition_penalty: 1.0
    frequency_penalty: 0.0
    presence_penalty: 0.0
    seed: 42
    top_k: 0
    min_p: 0.0
    top_p: 0.9

[ashim-mahara]

Hey ! Thanks for raising this issue, the --overide-batch-size issue should be fixed.

For adding more doc for adapter models can you provide me with an example model (and base model) on the hub you would like to see work ?

Thank you for the prompt reply and I'm sorry this got buried in the emails.

I am using a custom model that I finetuned. Uhh, but you could test it out with any PEFT model. I might release the model in the coming weeks - if I do I will post it here for good measure.

[NathanHB]

@sciarrilli yes because there is a missmatch between main and latest release, you would need to install from main to be able to use the examples

[NathanHB]

@ashim-mahara unfirtunatly i don't have the bandwidth right now.
To make it wor youwould need to install from main and edit this and here too part of the code:

If you open a PR will be happy to review :)