lora config

#6
by SeanLee97 - opened

Excellent work!

I've read the technical paper. It mentions the inclusion of LoRA adapters in all linear layers.
Based on my understanding, you specified the target_modules as follows: ['v_proj', 'q_proj', 'down_proj', 'k_proj', 'gate_proj', 'o_proj', 'up_proj'].
Is my understanding correct?

Could you share the detailed Lora config? Thx!

Owner

Thanks @SeanLee97 ! I learn a lot from your AnglE paper!

The released model contains merged lora weights, I thought this would make it more convenient to use without depending on PEFT library.

Here is our LoRA config:

{
  "auto_mapping": null,
  "base_model_name_or_path": "mistralai/Mistral-7B-v0.1",
  "bias": "none",
  "fan_in_fan_out": false,
  "inference_mode": true,
  "init_lora_weights": true,
  "layers_pattern": null,
  "layers_to_transform": null,
  "lora_alpha": 32,
  "lora_dropout": 0.1,
  "modules_to_save": null,
  "peft_type": "LORA",
  "r": 16,
  "revision": null,
  "target_modules": [
    "q_proj",
    "k_proj",
    "v_proj",
    "o_proj",
    "down_proj",
    "up_proj",
    "gate_proj"
  ],
  "task_type": "FEATURE_EXTRACTION"
}

Sign up or log in to comment