Upload 5 files

README.md

@@ -0,0 +1,3 @@
+# Milo X1 (dpo)
+
+Transformer en español entrenado desde cero. ~12.3M parámetros, GQA + RoPE + QK-Norm + SwiGLU + RMSNorm. Cargar con `milo_core.load_model`. Plantilla de chat con tokens especiales (`<|user|>`, `<|assistant|>`, `<|end|>`).

config.json

@@ -0,0 +1,37 @@
+{
+  "vocab_size": 16384,
+  "d_model": 320,
+  "n_layers": 6,
+  "n_heads": 5,
+  "n_kv_heads": 1,
+  "d_ff": 960,
+  "context_length": 768,
+  "dropout": 0.0,
+  "rope_theta": 100000.0,
+  "qk_norm": true,
+  "logit_softcap": 30.0,
+  "z_loss": 0.0001,
+  "tie_embeddings": true,
+  "architecture": "MiloX1",
+  "stage": "dpo",
+  "special_tokens": [
+    "<PAD>",
+    "<UNK>",
+    "<BOS>",
+    "<EOS>",
+    "<|system|>",
+    "<|user|>",
+    "<|assistant|>",
+    "<|end|>",
+    "<|reserved_0|>",
+    "<|reserved_1|>",
+    "<|reserved_2|>",
+    "<|reserved_3|>",
+    "<|reserved_4|>",
+    "<|reserved_5|>",
+    "<|reserved_6|>",
+    "<|reserved_7|>"
+  ],
+  "chat_template": "<BOS>[<|system|>S<|end|>]<|user|>U<|end|><|assistant|>A<|end|>",
+  "core_version": "4.0"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c5a8585dbbc84fd5c63910330f9990240707f529a39dd5e87cc55703e68a72ed
+size 49013720