Add a response template and generation tags

chat_template.jinja

@@ -111,9 +111,10 @@
                 {%- set content = content -%}
             {%- endif -%}
         {%- endif %}
+        {{- '<|im_start|>assistant\n' }}
+        {%- set turn %}
         {%- if message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
             {# Assistant message has tool calls. #}
-            {{- '<|im_start|>assistant\n' }}
                 {%- set include_content = not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
                 {%- if content is string and content | trim | length > 0 %}
                     {%- if include_content %}
@@ -149,11 +150,11 @@
                         {%- endif %}
                     {{- '</function>\n</tool_call>\n' -}}
                 {%- endfor %}
-                {{- '<|im_end|>\n' }}
+                {{- '<|im_end|>' }}
         {%- else %}
             {# Assistant message doesn't have tool calls. #}
             {%- if not (truncate_history_thinking and loop.index0 < ns.last_user_idx) %}
-                {{- '<|im_start|>assistant\n' ~ (content | default('', true) | string | trim) ~ '<|im_end|>\n' }}
+                {{- (content | default('', true) | string | trim) ~ '<|im_end|>' }}
             {%- else %}
                 {%- set c = (content | default('', true) | string) %}
                 {%- if '<think>' in c and '</think>' in c %}
@@ -161,12 +162,33 @@
                 {%- endif %}
                 {%- set c = c | trim %}
                 {%- if c | length > 0 %}
-                    {{- '<|im_start|>assistant\n' ~ c ~ '<|im_end|>\n' }}
+                    {{- c ~ '<|im_end|>' }}
                 {%- else %}
-                    {{- '<|im_start|>assistant\n<|im_end|>\n' }}
+                    {{- '<|im_end|>' }}
                 {%- endif %}
             {%- endif %}
         {%- endif %}
+        {%- endset %}
+        {#- What the generation prompt pre-fills stays out of the generation block: the model never generates it #}
+        {%- set prefill = '<think>\n' if enable_thinking else '<think></think>' %}
+        {%- if turn.startswith(prefill) %}
+            {%- set lead = prefill %}
+        {%- elif not turn.startswith('<think>') %}
+            {%- set lead = '' %}
+        {%- elif enable_thinking or '</think>' not in turn %}
+            {%- set lead = '<think>' %}
+        {%- else %}
+            {#- Reasoning with thinking off: the prompt pre-fills an empty block, so the whole block stays out #}
+            {%- set lead = turn.split('</think>')[0] ~ '</think>' %}
+            {%- if turn[lead | length:].startswith('\n') %}
+                {%- set lead = lead ~ '\n' %}
+            {%- endif %}
+        {%- endif %}
+        {{- lead }}
+        {%- generation %}
+        {{- turn[lead | length:] }}
+        {%- endgeneration %}
+        {{- '\n' }}
     {%- elif message.role == "user" or message.role == "system" %}
         {{- '<|im_start|>' + message.role + '\n' }}
         {%- set content = message.content | string %}

tokenizer_config.json

@@ -8014,5 +8014,43 @@
   ],
   "model_max_length": 262144,
   "tokenizer_class": "PreTrainedTokenizerFast",
-  "unk_token": "<unk>"
+  "unk_token": "<unk>",
+  "response_template": {
+    "defaults": {
+      "role": "assistant"
+    },
+    "start_anchor": "<|im_start|>assistant\n",
+    "fields": {
+      "reasoning_content": {
+        "open": "<think>",
+        "close": "</think>",
+        "content": "text"
+      },
+      "tool_calls": {
+        "open_pattern": "<tool_call>\\s*<function=(?P<name>[^>\\n]+)>",
+        "close_pattern": "</tool_call>|(?=<tool_call>)",
+        "repeats": true,
+        "content": "xml-inline",
+        "content_args": {
+          "tag_pattern": "<parameter=(?P<key>[^>\\n]+)>\\n?(?P<value>.*?)\\n?(?:</parameter>|(?=<parameter=)|(?=</function>))"
+        },
+        "transform": {
+          "type": "function",
+          "function": {
+            "name": "{name}",
+            "arguments": "{content}"
+          }
+        }
+      },
+      "content": {
+        "close": [
+          "<|im_end|>",
+          "<|endoftext|>"
+        ],
+        "content": "text",
+        "repeats": true,
+        "join": ""
+      }
+    }
+  }
 }