sft_with_kl_qwen-4B_NR-short-32k-16-1k-8_lr-3e-06-constant-bs-512_steps-128/chat_template.jinja

{# Very similar to sft-with-offline-thoughts, but the thought is now included in the "suffix" #}
{% if messages[0]['role'] == 'prompt' -%}
<|im_start|>system
You are a concise problem-solving assistant.

For each problem:
- Think inside <thinking> and </thinking> tags.
- Keep thinking under 500 words total — be concise, no filler.
- Make absolutely sure to end your thinking with </thinking>.
- After thinking, write the final result inside <answer> and </answer> tags.
- Do not include any text outside these tags

<|im_start|>user
Problem:
{{ messages[0]['content'].strip() }}<|im_end|>
<|im_start|>assistant
<thinking>

{%- elif messages[0]['role'] == 'thought' -%}
{{ messages[0]['content'].strip() }}
{%- elif messages[0]['role'] == 'suffix' -%}
<answer>
{{ messages[0]['content'].strip() }}
</answer><|im_end|><|endoftext|>
{%- endif %}