Upload 6 files

訓練 tokenizer 增加 Python 程式碼資料集以及維基文庫，訓練時增加維基文庫。模型學會了古文、古詩詞的語法。

config.json

@@ -7,7 +7,7 @@
     "AutoConfig": "modeling_kuixing.KuiXingHFConfig",
     "AutoModelForCausalLM": "modeling_kuixing.KuiXingForCausalLM"
   },
-  "vocab_size": 109568,
+  "vocab_size": 112640,
   "hidden_size": 2400,
   "num_hidden_layers": 16,
   "num_attention_heads": 32,
@@ -18,5 +18,6 @@
   "pad_token_id": 0,
   "bos_token_id": 2,
   "eos_token_id": 3,
-  "torch_dtype": "float32"
+  "torch_dtype": "float32",
+  "transformers_version": "5.9.0"
 }

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cc60ca46ed58edc2644a2b66a931dd28ced77b4df94b9fea62369cc9bbd46572
-size 6537547232
+oid sha256:d3c1c605200c53792654eec2ac99ec76cac9b2387ef664c3e3e3876981269da5
+size 5515185528

modeling_kuixing.py

@@ -19,7 +19,7 @@ class KuiXingHFConfig(PretrainedConfig):
     model_type = "kuixing"
     def __init__(
         self,
-        vocab_size=109568,
+        vocab_size=112640,
         hidden_size=2400,
         num_hidden_layers=16,
         num_attention_heads=32,
@@ -160,7 +160,7 @@ class KuiXingForCausalLM(PreTrainedModel, GenerationMixin):
         # lm_head.weight 是刻意省略的（weight tying），從 missing 中排除再判斷
         missing = [k for k in missing if k != "model.lm_head.weight"]
         if not missing and not unexpected:
-            print("✅ 所有權重 key 完整對映，無缺漏。")
+            print("✅ 魁星所有權重 key 完整對映，無缺漏。")
         else:
             if missing:    print(f"⚠️  缺少 key（{len(missing)}）：{missing[:5]}")
             if unexpected: print(f"⚠️  多餘 key（{len(unexpected)}）：{unexpected[:5]}")

tokenizer.model

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ac0318ec811b0697f8ddffec0ef7b4603483f0c4fced21ba1c2a9b79e42c5662
-size 1621927
+oid sha256:13698a9916d5f488fcabeec6dc9bd6de65986ae171911ed6c5ab4af43b22c849
+size 1846838