Fix default temp=0.8, top_k=20, fix fallback tokenizer to 38 chars

server.py

@@ -264,7 +264,7 @@ vocab_size = hp["vocab_size"]
 
 # Fallback tokenizer if vocab.json missing
 if tok is None:
-    chars = sorted(set("abcdefghijklmnopqrstuvwxyz ."))
+    chars = [" ","!","\"","'","(",")",",","-",".",":",";","?","a","b","c","d","e","f","g","h","i","j","k","l","m","n","o","p","q","r","s","t","u","v","w","x","y","z"]
     tok = CharTokenizer(chars)
     print(f"  Built fallback char vocab: {tok.vocab_size} chars")
 
@@ -335,8 +335,8 @@ class ChatRequest(BaseModel):
     model: Optional[str] = MODEL_ID
     messages: List[Message]
     max_tokens: Optional[int] = 200
-    temperature: Optional[float] = 0.1
-    top_k: Optional[int] = 8
+    temperature: Optional[float] = 0.8
+    top_k: Optional[int] = 20
     repetition_penalty: Optional[float] = 1.3
     n: Optional[int] = 1
     stream: Optional[bool] = False
@@ -386,8 +386,8 @@ def chat_completions(req: ChatRequest):
         ids = [0]
 
     max_tokens = max(1, min(req.max_tokens or 200, block_size))
-    temperature = max(0.01, min(req.temperature or 0.1, 2.0))
-    top_k = max(1, min(req.top_k or 8, tok.vocab_size))
+    temperature = max(0.01, min(req.temperature or 0.8, 2.0))
+    top_k = max(1, min(req.top_k or 20, tok.vocab_size))
     rep_penalty = max(1.0, min(req.repetition_penalty or 1.3, 3.0))
     n = max(1, min(req.n or 1, 4))
     completion_id = f"chatcmpl-{uuid.uuid4().hex[:8]}"