From f46b995ea0fdecccff86df31de29d7ace715c11d Mon Sep 17 00:00:00 2001 From: markus583 Date: Tue, 21 May 2024 11:12:06 +0000 Subject: [PATCH] fix num layers --- configs/peft/lora_xlmr.json | 3 ++- wtpsplit/train/train_adapter.py | 2 +- 2 files changed, 3 insertions(+), 2 deletions(-) diff --git a/configs/peft/lora_xlmr.json b/configs/peft/lora_xlmr.json index 8b9a3ab3..a061210f 100644 --- a/configs/peft/lora_xlmr.json +++ b/configs/peft/lora_xlmr.json @@ -1,8 +1,9 @@ { "model_name_or_path": "xlm-roberta-base", - "output_dir": "xlmr-base-3_lora-v2_ep30_s10k", + "output_dir": "xlmr-base-3l_lora-v2_ep30_s10k", "block_size": 256, "eval_stride": 128, + "num_hidden_layers": 3, "do_train": true, "do_eval": true, "per_device_train_batch_size": 64, diff --git a/wtpsplit/train/train_adapter.py b/wtpsplit/train/train_adapter.py index 423e1ac6..bc38c30f 100644 --- a/wtpsplit/train/train_adapter.py +++ b/wtpsplit/train/train_adapter.py @@ -57,7 +57,7 @@ class Args: adapter_warmup_steps: int = 0 adapter_lr_multiplier: float = 1.0 text_column: str = "text" - num_hidden_layers = None + num_hidden_layers: int = 0 # NEW PARAMS use_subwords: bool = False