Best prompt and model

scripts/generate_scripts.py

@@ -13,7 +13,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 
 @spaces.GPU()
 def generate_local(prompt: str, max_new_tokens: int = 350, temperature: float = 0.7) -> str:
-    model_id = "Qwen/Qwen2.5-1.5B"
+    model_id = "Qwen/Qwen2.5-3B"
     device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # get the device the model is on
 
     tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
@@ -33,13 +33,16 @@ def generate_local(prompt: str, max_new_tokens: int = 350, temperature: float =
 
 def generate_script(prompt: str, word_count: int = 60) -> str:
     system_prompt = (
-        "You are a professional scriptwriter. "
-        "Generate the *exact spoken words* for a short narration in a YouTube video. "
-        f"The topic is: {prompt.strip()}. "
-        f"The script must be exactly {word_count} words long. "
-        "Do NOT include stage directions, scene descriptions, or formatting. "
-        "Only output the spoken words, as they should be said, in plain text. "
-        "Make it clear, engaging, and natural for a voiceover."
+        "You are an expert YouTube scriptwriter. "
+        "Your job is to write the EXACT words that will be spoken aloud in a video. "
+        f"Topic: {prompt.strip()}\n\n"
+        "🎯 Output rules:\n"
+        f"- Exactly {word_count} words.\n"
+        "- Only the spoken words. NO scene descriptions, instructions, or formatting.\n"
+        "- Write in natural, clear, and simple English, as if it's being said by a voiceover artist.\n"
+        "- Keep a steady rhythm (about 2 words per second).\n"
+        "- Do NOT include any explanations, labels, or headers. Only output the final spoken script.\n\n"
+        "Start now:"
     )
     return generate_local(system_prompt)