Upload 137 files

hugging/td_fuse/validate.py

@@ -180,6 +180,10 @@ def _format_chat_prompt(tokenizer, user_message: str, enable_thinking: bool = Tr
             add_generation_prompt=True,
             enable_thinking=enable_thinking,
         )
+        # Verify the template actually produced thinking tokens
+        if enable_thinking and "<think>" not in text:
+            # Template didn't add thinking trigger — use manual format
+            raise ValueError("Template missing think trigger")
         inputs = tokenizer(text, return_tensors="pt")
         return inputs
     except Exception:
@@ -187,7 +191,8 @@ def _format_chat_prompt(tokenizer, user_message: str, enable_thinking: bool = Tr
 
     # Fallback: manual Qwen3 chat format
     if enable_thinking:
-        text = f"<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant\n"
+        # Qwen3 thinking mode: start assistant turn with <think> to trigger CoT
+        text = f"<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant\n<think>\n"
     else:
         text = f"<|im_start|>user\n{user_message}<|im_end|>\n<|im_start|>assistant\n/no_think\n"
     inputs = tokenizer(text, return_tensors="pt")
@@ -222,15 +227,18 @@ def test_thinking_mode(
     new_tokens = outputs[0][inputs["input_ids"].shape[1]:]
     response = tokenizer.decode(new_tokens, skip_special_tokens=False)
 
-    # Check for thinking tags
-    has_think_open = "<think>" in response
+    # Check for thinking tags (we may have prefilled <think> in the prompt,
+    # so check for </think> which the model must produce to end thinking)
     has_think_close = "</think>" in response
-    passed = has_think_open and has_think_close
+    # If template handled it, <think> appears in new tokens too
+    has_think_open = "<think>" in response
+    # Pass if model produced </think> (thinking happened, whether <think> was prefilled or not)
+    passed = has_think_close
 
     print(f"\n[validate] Thinking mode test:")
     print(f"  Prompt:    {prompt}")
     print(f"  Response:  {response[:300]}...")
-    print(f"  <think>:   {'✓ found' if has_think_open else '✗ missing'}")
+    print(f"  <think>:   {'✓ found' if has_think_open else '(prefilled in prompt)'}")
     print(f"  </think>:  {'✓ found' if has_think_close else '✗ missing'}")
     print(f"  Status:    {'✓ PASS' if passed else '✗ FAIL'}")
 

hugging/td_lang/compiler.py

@@ -958,7 +958,7 @@ DO NOT EDIT - regenerate from the .td file instead.
         - Supports configurable sample count (cmd.n_samples if provided).
         - Produces domain-specific prompts (math, code, logic, factual).
         """
-        n_samples_expr = f"getattr(cmd, 'n_samples', 100)"  # static string for emit clarity
+        n_samples_val = getattr(cmd, 'n_samples', 100)  # resolve at compile time
         self._emit(f'print("[td_lang] Generating synthetic data for {cmd.target}...")')
         self._emit(f'checkpoint = models.get("{cmd.target}", {{}}).get("checkpoint")')
         self._emit("if not checkpoint:")
@@ -1092,7 +1092,7 @@ DO NOT EDIT - regenerate from the .td file instead.
         self._indent -= 1
         self._emit("")
         self._emit("synth_data = []")
-        self._emit(f"n_samples = getattr(cmd, 'n_samples', 100)")
+        self._emit(f"n_samples = {n_samples_val}")
         self._emit("for i in range(n_samples):")
         self._indent += 1
         self._emit("domain = random.choice(weak_topics)")