fix: enable sampling — temperature=0.8 for MCQ, 0.4 for evaluator

Greedy decoding (do_sample=False) caused the model to repeat the same
answer for all 4 MCQ options (e.g. 4x BERT) and to copy source text
verbatim for the model answer. Sampling breaks the repetition loop and
produces diverse, coherent outputs.

core/evaluator.py

@@ -50,4 +50,4 @@ def evaluate_answer(question: str, chunk: str, student_answer: str, language: st
     )
     # 4-section feedback fits comfortably in 320 tokens — keeps CPU
     # (llama.cpp) latency inside the UI timeout.
-    return llm.generate(prompt, max_new_tokens=320).strip()
+    return llm.generate(prompt, max_new_tokens=320, temperature=0.4).strip()

core/questioner.py

@@ -96,7 +96,7 @@ def generate_mcq(chunk: str, language: str = "English") -> dict:
     prompt = _MCQ_TEMPLATE.format(chunk=chunk.strip(), language=language)
     mcq: dict = {}
     for _ in range(3):
-        raw = llm.generate(prompt).strip()
+        raw = llm.generate(prompt, temperature=0.8).strip()
         mcq = parse_mcq(raw)
         choices = list(mcq.get("choices", {}).values())
         if len(choices) == 4 and len({c.lower().strip() for c in choices}) == 4:

model/llm.py

@@ -90,17 +90,20 @@ class LLM:
             tokenizer=self._tokenizer,
         )
 
-    def generate(self, prompt: str, max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS) -> str:
+    def generate(self, prompt: str, max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS, temperature: float = 0.0) -> str:
         """Run *prompt* through the model and return the generated text only."""
         messages = [{"role": "user", "content": prompt}]
         text = self._tokenizer.apply_chat_template(
             messages, tokenize=False, add_generation_prompt=True,
             enable_thinking=False,
         )
+        sample = temperature > 0.0
         output = self._pipe(
             text,
             max_new_tokens=max_new_tokens,
-            do_sample=False,
+            do_sample=sample,
+            temperature=temperature if sample else None,
+            top_p=0.95 if sample else None,
             return_full_text=False,
         )
         return output[0]["generated_text"]
@@ -131,11 +134,11 @@ class LlamaCppLLM:
             verbose=False,
         )
 
-    def generate(self, prompt: str, max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS) -> str:
+    def generate(self, prompt: str, max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS, temperature: float = 0.0) -> str:
         out = self._llm.create_chat_completion(
             messages=[{"role": "user", "content": prompt}],
             max_tokens=max_new_tokens,
-            temperature=0.0,
+            temperature=temperature,
         )
         return out["choices"][0]["message"]["content"]