docs: add a fine tune samples

Author: phodal

README.md

@@ -851,6 +851,65 @@ if __name__ == "__main__":
 
 ![Finetune Model Choice](images/finetune-model-choice.jpg)
 
+#### 数据集信息
+
+由 Unit Eval + OSS Instruct 数据集构建而来：
+
+- 3000 条补全（Inline，InBlock，AfterBlock）数据集。
+- 1500 条单元测试数据集。
+- 4000 条 OSS Instruct 数据集。
+
+#### 参数示例：
+
+```bash
+!cd DeepSeek-Coder/finetune && deepspeed finetune_deepseekcoder.py \
+    --model_name_or_path $MODEL_PATH \
+    --data_path $DATA_PATH \
+    --output_dir $OUTPUT_PATH \
+    --num_train_epochs 1 \
+    --model_max_length 1024 \
+    --per_device_train_batch_size 2 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 1 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 2000 \
+    --save_total_limit 10 \
+    --learning_rate 1e-4 \
+    --warmup_steps 10 \
+    --logging_steps 1 \
+    --lr_scheduler_type "cosine" \
+    --gradient_checkpointing True \
+    --report_to "tensorboard" \
+    --deepspeed configs/ds_config_zero3.json \
+    --bf16 True
+```
+
+运行日志：
+
+```bash
+`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...
+  0%|                                                  | 0/2125 [00:00<?, ?it/s]`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...
+{'loss': 3.9356, 'learning_rate': 0.0, 'epoch': 0.0}                            
+{'loss': 0.8462, 'learning_rate': 3.0102999566398115e-05, 'epoch': 0.0}         
+{'loss': 0.909, 'learning_rate': 4.771212547196624e-05, 'epoch': 0.0}           
+{'loss': 0.3674, 'learning_rate': 6.020599913279623e-05, 'epoch': 0.0}          
+{'loss': 0.3959, 'learning_rate': 6.989700043360187e-05, 'epoch': 0.0}          
+{'loss': 0.7964, 'learning_rate': 7.781512503836436e-05, 'epoch': 0.0}          
+{'loss': 0.3542, 'learning_rate': 8.450980400142567e-05, 'epoch': 0.0}          
+{'loss': 1.7094, 'learning_rate': 9.030899869919434e-05, 'epoch': 0.0}          
+{'loss': 0.5968, 'learning_rate': 9.542425094393248e-05, 'epoch': 0.0}          
+{'loss': 0.6208, 'learning_rate': 9.999999999999999e-05, 'epoch': 0.0}          
+{'loss': 0.4074, 'learning_rate': 0.0001, 'epoch': 0.01}                        
+{'loss': 0.3637, 'learning_rate': 0.0001, 'epoch': 0.01}                        
+{'loss': 0.3459, 'learning_rate': 0.0001, 'epoch': 0.01}                        
+{'loss': 0.6971, 'learning_rate': 0.0001, 'epoch': 0.01}                        
+{'loss': 0.3917, 'learning_rate': 0.0001, 'epoch': 0.01}                        
+{'loss': 0.5859, 'learning_rate': 0.0001, 'epoch': 0.01}                        
+{'loss': 0.5923, 'learning_rate': 0.0001, 'epoch': 0.01}                        
+  1%|▎                                     | 17/2125 [05:14<10:03:38, 17.18s/it]
+```
+
 其它：
 
 - 详细的 Notebook 见：[code/finetune/finetune.ipynb](code/finetune/finetune.ipynb)