换用github jwjohns/unsloth-GRPO-qwen2.5 验证GRPO训练模型

src/qwen_notebook_clone.py

@@ -17,7 +17,7 @@ PatchFastRL("GRPO", FastLanguageModel)
 model_name = f"../models/pretrained/DeepSeek-R1-Distill-Qwen-1.5B"
 model, tokenizer = FastLanguageModel.from_pretrained(
     model_name=model_name,
-    max_seq_length=512, # 2028
+    max_seq_length=2048, # 2048
     load_in_4bit=True,
     fast_inference=False,
     max_lora_rank=128,
@@ -151,8 +151,8 @@ training_args = GRPOConfig(
     num_generations = 8,
     max_prompt_length = 256,
     max_completion_length = 200,
-    max_steps = 2000,  # Increased 8x for longer training
-    save_steps = 500,  # Save checkpoints more frequently
+    max_steps = 250,  # 2000 Increased 8x for longer training
+    save_steps = 250,  # Save checkpoints more frequently
     max_grad_norm = 0.1,
     report_to = "tensorboard",  # Enable tensorboard reporting for metrics display
     output_dir = "outputs",