{#- ChatML template with assistant-generation markers (transformers convention):
    wrapping the assistant turn in {% generation %}...{% endgeneration %} lets
    apply_chat_template(..., return_assistant_tokens_mask=True) mark the assistant
    tokens so assistant_only_loss masks the prompt. Assistant content is passed
    through verbatim, so the <think>...</think> trace in the data is preserved.
    A math system prompt is injected (own wording); a caller-supplied system
    message, if any, is appended after it. #}
{%- set enable_thinking = true -%}
{%- set sys_instruction = "You are a mathematics problem solver. Work through each problem with brief, focused reasoning: show only the steps that matter, keeping it short when a short path suffices and taking more space only when the problem genuinely needs it. Give the final answer exactly once, enclosed in \\boxed{...}, as the final value in exact form. Put nothing else inside the box (no words, units, or working), and write nothing after it." -%}
{%- if messages|length > 0 and messages[0]['role'] == 'system' %}
{{- '<|im_start|>system\n' + sys_instruction + '\n\n' + messages[0]['content'] + '<|im_end|>\n' }}
{%- set loop_messages = messages[1:] %}
{%- else %}
{{- '<|im_start|>system\n' + sys_instruction + '<|im_end|>\n' }}
{%- set loop_messages = messages %}
{%- endif %}
{%- for message in loop_messages %}
{%- if message['role'] == 'assistant' %}
{{- '<|im_start|>assistant\n' }}{% generation %}{{ message['content'] }}{{ '<|im_end|>' }}{% endgeneration %}{{ '\n' }}
{%- else %}
{{- '<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>\n' }}
{%- endif %}
{%- endfor %}
{%- if add_generation_prompt %}
{{- '<|im_start|>assistant\n' }}
{%- endif %}