From beb97a099c57121edaf1eae28bb949c8b996ec14 Mon Sep 17 00:00:00 2001
From: wql <wengqinlan18@163.com>
Date: Tue, 13 Aug 2024 13:19:42 +0800
Subject: [PATCH] train: change yaml

---
 examples/train_lora/llama2_lora_sft.yaml | 27 ++++++++++++------------
 1 file changed, 14 insertions(+), 13 deletions(-)

diff --git a/examples/train_lora/llama2_lora_sft.yaml b/examples/train_lora/llama2_lora_sft.yaml
index 0e0e5fef..95884399 100644
--- a/examples/train_lora/llama2_lora_sft.yaml
+++ b/examples/train_lora/llama2_lora_sft.yaml
@@ -5,17 +5,18 @@ model_name_or_path: modelscope/Llama-2-7b-ms
 stage: sft
 do_train: true
 finetuning_type: lora
-lora_target: q_proj,v_proj
+lora_target: all
 
 ### dataset
 dataset: alpaca_en
 template: llama2
 cutoff_len: 1024
-max_samples: 100000
+max_samples: 10000
 overwrite_cache: true
+preprocessing_num_workers: 16
 
 ### output
-output_dir: ./saves/LLaMA2-7B/lora/train_24_8_13_10_02
+output_dir: ./saves/LLaMA2-7B/lora/train_24_8_13_13_16
 logging_steps: 3
 save_steps: 100
 plot_loss: true
@@ -24,15 +25,15 @@ overwrite_output_dir: true
 ### train
 per_device_train_batch_size: 2
 gradient_accumulation_steps: 8
-learning_rate: 5.0e-05
-num_train_epochs: 10
+learning_rate: 1.0e-4
+num_train_epochs: 10.0
 lr_scheduler_type: cosine
-fp16: true
+warmup_ratio: 0.1
+bf16: true
+ddp_timeout: 180000000
 
-max_grad_norm: 1.0
-warmup_steps: 0
-optim: adamw_torch
-report_to: none
-lora_rank: 8
-lora_alpha: 16
-lora_dropout: 0.1
\ No newline at end of file
+### eval
+val_size: 0.1
+per_device_eval_batch_size: 2
+eval_strategy: steps
+eval_steps: 500
\ No newline at end of file