add trust_remote_code, explicit bfloat16 and bnb dtype settings

human_chat_qlora.yml

@@ -18,6 +18,13 @@ tokenizer_type: AutoTokenizer
 
 load_in_4bit: true
 strict: false
+trust_remote_code: true
+
+# Explicit dtype settings to avoid weight conversion errors
+torch_dtype: bfloat16
+bnb_4bit_use_double_quant: true
+bnb_4bit_quant_type: nf4
+bnb_4bit_compute_dtype: bfloat16
 
 # --- System prompt baked into every conversation ---
 # This is the primary lever for "no error correcting, more human-like"