Upload folder using huggingface_hub

convert_checkpoint.py

@@ -0,0 +1,54 @@
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import os
+
+# --- 1. 配置你的路径 ---
+
+# 你微调时使用的“基础模型”的Hugging Face Hub ID
+# ！！！这一步至关重要，必须是你开始训练时用的那个模型！！！
+base_model_id = "Qwen/Qwen1.5-7B" 
+
+# 你保存的训练检查点 .pth 文件的路径
+checkpoint_path = "/path/to/your/checkpoint.pth"  # <--- 修改这里：你的.pth文件路径
+
+# 你想将转换后的、可供评测的模型保存到哪里？
+# 这会是一个新的目录，lm-evaluation-harness将从这里加载
+output_dir = "../my_pretrain_for_eval" # <--- 修改这里：定义你的输出目录名
+
+# --- 2. 加载基础模型和Tokenizer ---
+# 这一步是获取模型的“骨架”（架构）和分词器
+
+print(f"Loading base model architecture from: {base_model_id}")
+# trust_remote_code=True 对于Qwen等模型是必需的
+model = AutoModelForCausalLM.from_pretrained(base_model_id, trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)
+
+# --- 3. 加载你的检查点并提取模型权重 ---
+
+print(f"Loading training checkpoint from: {checkpoint_path}")
+# 加载你保存的字典
+checkpoint = torch.load(checkpoint_path, map_location='cpu') # 使用map_location='cpu'以避免GPU内存问题
+
+# 从字典中提取模型的 state_dict
+# 根据你保存的代码，权重在 'model' 这个键下
+model_weights = checkpoint['model']
+
+# --- 4. 将你的权重加载到模型“骨架”中 ---
+
+print("Loading fine-tuned weights into the model...")
+model.load_state_dict(model_weights)
+print("Weights loaded successfully.")
+
+# --- 5. 将模型和Tokenizer保存为标准的Hugging Face格式 ---
+
+print(f"Saving model and tokenizer to: {output_dir}")
+os.makedirs(output_dir, exist_ok=True)
+
+# model.save_pretrained() 会自动创建 config.json, pytorch_model.bin 等文件
+model.save_pretrained(output_dir)
+
+# tokenizer.save_pretrained() 会自动创建 tokenizer.json 等文件
+tokenizer.save_pretrained(output_dir)
+
+print("Conversion complete!")
+print(f"Your model is now ready for evaluation at: {output_dir}")

latest_checkpoint.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:875067c4a946633699e0f8875a243032d9abfdfe7b25463b52adb735d7501d37
+size 310056955