!27 added example yaml files for qwen2.5 distillation and deepseek-r1-distill-qwen sft

Merge pull request !27 from A1waysBeenHere/master

examples/deepseek_r1/train_sft_full_deepseek_r1.yaml

@@ -0,0 +1,28 @@
+# model
+model_id: DeepSeek-r1-distill-qwen-7b
+
+# method
+stage: sft
+do_train: true
+finetuning_type: full
+deepspeed: examples/deepspeed/ds_z2_config.json
+
+# dataset
+dataset: alpaca_zh_51k
+cutoff_len: 1024
+template: deepseek_r1
+
+# output
+output_dir: saves/deepseek_r1_distill_qwen_7b_lora
+logging_steps: 1
+save_steps: 5000
+overwrite_output_dir: true
+
+# train
+per_device_train_batch_size: 8
+learning_rate: 1.0e-5
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+max_steps: 5000
+seed: 1234

examples/deepseek_r1/train_sft_lora_deepseek_r1.yaml

@@ -15,7 +15,7 @@ template: deepseek_r1
 # output
 output_dir: saves/deepseek_r1_distill_qwen_7b_lora
 logging_steps: 1
-save_steps: 500
+save_steps: 5000
 overwrite_output_dir: true
 
 # train
@@ -24,5 +24,5 @@ learning_rate: 1.0e-5
 lr_scheduler_type: cosine
 warmup_ratio: 0.1
 bf16: true
-max_steps: 2000
+max_steps: 5000
 seed: 1234

examples/qwen2.5/train_distill_full_qwen2_5_7b_instruct.yaml

@@ -13,7 +13,7 @@ cutoff_len: 1024
 template: deepseek_r1
 
 # output
-output_dir: saves/qwen2.5_7b_chat_lora
+output_dir: saves/qwen2.5_7b_chat_distill_full
 logging_steps: 1
 save_steps: 20000
 overwrite_output_dir: true

examples/qwen2.5/train_distill_lora_qwen2_5_7b_instruct.yaml

@@ -0,0 +1,28 @@
+# model
+model_id: Qwen2.5-7B-Chat
+
+# method
+stage: sft
+do_train: true
+finetuning_type: lora
+deepspeed: examples/deepspeed/ds_z2_config.json
+
+# dataset
+dataset: sky_t1_17k
+cutoff_len: 1024
+template: deepseek_r1
+
+# output
+output_dir: saves/qwen2.5_7b_chat_distill_lora
+logging_steps: 1
+save_steps: 20000
+overwrite_output_dir: true
+
+# train
+per_device_train_batch_size: 8
+learning_rate: 1.0e-5
+lr_scheduler_type: cosine
+warmup_ratio: 0.1
+bf16: true
+max_steps: 5000
+seed: 1234