add train sft examples (#2491)

Co-authored-by: llbdyiu66 <[email protected]>

Author: llbdyiu66

examples/README.md

@@ -0,0 +1,43 @@
+## 精调
+
+### 数据准备
+
+我们支持的精调数据格式是每行包含一个字典的 json 文件，每个字典包含以下字段：
+
+- `src` : `str, List(str)`, 模型的输入指令（instruction）、提示（prompt），模型应该执行的任务。
+- `tgt` : `str, List(str)`, 模型的输出。
+
+样例数据：
+
+```text
+{"src": "Give three tips for staying healthy.", "tgt": "1.Eat a balanced diet and make sure to include plenty of fruits and vegetables. \n2. Exercise regularly to keep your body active and strong. \n3. Get enough sleep and maintain a consistent sleep schedule."}
+...
+```
+
+为了方便测试，我们也提供了[tatsu-lab/alpaca](https://huggingface.co/datasets/tatsu-lab/alpaca)demo 数据集可以直接使用：
+
+```shell
+wget https://bj.bcebos.com/paddlenlp/datasets/examples/alpaca_demo.gz
+tar -xvf alpaca_demo.gz
+```
+
+### 全参精调：SFT
+
+单卡
+```bash
+# 需要12G显存左右
+python -u run_finetune.py ./config/qwen/sft_argument_qwen2_0p5b.json
+```
+
+多卡
+```bash
+python -u -m paddle.distributed.launch --devices "0,1,2,3,4,5,6,7" run_finetune.py ./config/qwen/sft_argument_qwen2_0p5b.json
+```
+
+### LoRA
+
+LoRA 启动命令参考
+```bash
+# 需要9G左右显存
+python -u run_finetune.py ./config/qwen/lora_argument_qwen2_0p5b.json
+```

examples/config/qwen/lora_argument_qwen2_0p5b.json

@@ -0,0 +1,36 @@
+{
+    "model_name_or_path": "PaddleNLP/Qwen2-0.5B-Instruct",
+    "dataset_name_or_path": "./data/sft",
+    "output_dir": "./checkpoints/qwen2_paddle_lora_ckpts",
+    "per_device_train_batch_size": 1,
+    "gradient_accumulation_steps": 4,
+    "per_device_eval_batch_size": 8,
+    "eval_accumulation_steps":16,
+    "num_train_epochs": 1,
+    "learning_rate": 3e-05,
+    "warmup_steps": 10,
+    "logging_steps": 1,
+    "evaluation_strategy": "epoch",
+    "save_strategy": "epoch",
+    "src_length": 1024,
+    "max_length": 2048,
+    "fp16": true,
+    "fp16_opt_level": "O2",
+    "do_train": true,
+    "do_eval": true,
+    "disable_tqdm": true,
+    "load_best_model_at_end": true,
+    "eval_with_do_generation": false,
+    "metric_for_best_model": "accuracy",
+    "recompute": true,
+    "save_total_limit": 1,
+    "tensor_parallel_degree": 1,
+    "pipeline_parallel_degree": 1,
+    "sharding": "stage2",
+    "lora": true,
+    "zero_padding": false,
+    "unified_checkpoint": true,
+    "use_flash_attention": false,
+    "pissa": false,
+    "use_mora": false
+  }

examples/config/qwen/sft_argument_qwen2_0p5b.json

@@ -0,0 +1,33 @@
+{
+    "model_name_or_path": "PaddleNLP/Qwen2-0.5B-Instruct",
+    "dataset_name_or_path": "./data/sft",
+    "output_dir": "./checkpoints/qwen2_paddle_sft_ckpts",
+    "per_device_train_batch_size": 1,
+    "gradient_accumulation_steps": 4,
+    "per_device_eval_batch_size": 8,
+    "eval_accumulation_steps":16,
+    "num_train_epochs": 1,
+    "learning_rate": 3e-05,
+    "warmup_steps": 10,
+    "logging_steps": 1,
+    "evaluation_strategy": "epoch",
+    "save_strategy": "epoch",
+    "src_length": 1024,
+    "max_length": 2048,
+    "fp16": true,
+    "fp16_opt_level": "O2",
+    "do_train": true,
+    "do_eval": true,
+    "disable_tqdm": true,
+    "load_best_model_at_end": true,
+    "eval_with_do_generation": false,
+    "metric_for_best_model": "accuracy",
+    "recompute": true,
+    "save_total_limit": 1,
+    "tensor_parallel_degree": 1,
+    "pipeline_parallel_degree": 1,
+    "sharding": "stage2",
+    "zero_padding": false,
+    "unified_checkpoint": true,
+    "use_flash_attention": false
+  }