Chat Templates

A chat template is a Jinja2 snippet that formats messages into the string a model was trained on. For example:

>>> from transformers import AutoTokenizer
>>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-0.5B-Instruct")
>>> tokenizer.chat_template
"{% for message in messages %}{% if loop.first and messages[0]['role'] != 'system' %}{{ '<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n' }}{% endif %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
>>> tokenizer.apply_chat_template([{"role": "user", "content": "Hi!"}], tokenize=False)
'<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n<|im_start|>user\nHi!<|im_end|>\n'

In most cases you don’t need to worry about chat templates: models ship their template along with the tokenizer, and TRL applies it for you. The whole thing is transparent. But some TRL recipes rely on features that most shipped templates don’t include:

• SFT with assistant_only_loss=True needs {% generation %} / {% endgeneration %} markers around assistant output, so the loss mask can target only assistant tokens.
• GRPO with tool calls needs the template to be prefix-preserving: appending a tool message must not change how earlier messages are rendered.

TRL ships patched templates under trl/chat_templates/ for common families (Qwen, Llama, DeepSeek-V3, GPT-OSS, …) and swaps them in automatically for supported models. For any other model, you’ll need to patch its template yourself. The rest of this page catalogs what’s bundled.

Supported model families

TRL stores reference copies of the original templates so it can identify supported models at init and swap in a training template when needed. The following families are recognized: Cohere, Cohere2, DeepSeek-V3 (including DeepSeek-R1, DeepSeek-R1-0528 and DeepSeek-V3-0324), DeepSeek-R1-Distill, DiffusionGemma, Gemma, Gemma3, Gemma4 (including Gemma-4-12B, Gemma-4-26B-A4B and Gemma-4-31B), GLM-4-MoE (4.5, 4.6, 4.7), GLM-5, GLM-5.1, GLM-5.2, GLM-5.3, GLM-5.3-Flash, GPT-OSS, Idefics3, LFM2, LFM2.5 (including LFM2.5-1.2B-Instruct, LFM2.5-2.6B and LFM2.5-24B-A2B), LFM2.5-VL (including LFM2.5-VL-450M), Llama 3 / 3.1 / 3.2, Llava-Next, Mistral Large 4, Mistral Medium 3.5, Muse Glimmer, Muse Spark Open, Nemotron 3 (Nano, Super, Ultra), Nemotron 3.5 Lightning, Phi-3, Phi-3.5, Qwen2-VL, Qwen2.5 (including the base checkpoints), Qwen2.5-VL, Qwen3 (including the Instruct-2507 and Thinking-2507 variants), Qwen3-Coder (including Qwen3-Coder-Next), Qwen3-VL, Qwen3.5, Qwen3.6, Qwen3.8, SmolVLM (including SmolVLM2).

Training templates

Patched templates that fix training-specific issues. Swapped in at init when tools are enabled (GRPO) or when assistant_only_loss=True (SFT).

cohere_training.jinja

Patched Cohere template. Diff vs cohere.jinja:

Wrap assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

cohere2_training.jinja

Patched Cohere2 template. Diff vs cohere2.jinja:

Move the trailing <|END_OF_TURN_TOKEN|> from after the role-dispatch {% endif %} into each role branch, and wrap the assistant branch (<|START_RESPONSE|>...<|END_RESPONSE|><|END_OF_TURN_TOKEN|>) with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

deepseek_r1_distill_training.jinja

Patched DeepSeek-R1-Distill template. Diff vs deepseek_r1_distill.jinja:

• Removes the '</think>' split on assistant content. The original drops everything up to and including </think>, which would silently remove the reasoning from the training target.
• Wraps assistant message output with {% generation %} / {% endgeneration %} markers for SFT assistant-only loss. The <｜Assistant｜> and <｜tool▁outputs▁end｜> that open an assistant turn stay outside the markers, since the prompt already ends with them.

deepseekv3_training.jinja

Patched DeepSeek-V3 template. Diff vs deepseekv3.jinja:

• Uses | tojson on tool['function']['arguments'] so that arguments can be passed as a dict (the documented format per transformers docs). The original template uses raw string concatenation, which crashes on dict inputs.
• Removes the '</think>' split on assistant content, so the reasoning block is always rendered. The original drops everything up to and including </think>, which would silently remove the reasoning from the training target.
• Wraps assistant message output with {% generation %} / {% endgeneration %} markers for SFT assistant-only loss. The <｜Assistant｜> and <｜tool▁outputs▁end｜> that open an assistant turn stay outside the markers, since the prompt already ends with them.

deepseekv3_v2_training.jinja

Patched DeepSeek-R1-0528 template. Diff vs deepseekv3_v2.jinja:

• Uses | tojson on tool['function']['arguments'] so that arguments can be passed as a dict, as in deepseekv3_training.jinja.
• Removes the '</think>' split on assistant content, so the reasoning block is always rendered. The original drops everything up to and including </think>, which would silently remove the reasoning from the training target.
• Wraps assistant message output with {% generation %} / {% endgeneration %} markers for SFT assistant-only loss. The <｜tool▁outputs▁end｜> that opens an assistant turn after tool outputs stays outside the markers, since the prompt already ends with it.

deepseekv3_v3_training.jinja

Patched DeepSeek-V3 template (used by deepseek-ai/DeepSeek-V3). Same diff as deepseek_r1_distill_training.jinja, applied to deepseekv3_v3.jinja: {% generation %} / {% endgeneration %} markers around assistant message output for SFT assistant-only loss, with the <｜Assistant｜> and <｜tool▁outputs▁end｜> that open an assistant turn left outside them.

deepseekv3_v4_training.jinja

Patched DeepSeek-V3-0324 template (used by deepseek-ai/DeepSeek-V3-0324). Same diff as deepseekv3_v2_training.jinja, applied to deepseekv3_v4.jinja: | tojson on tool-call arguments, and {% generation %} / {% endgeneration %} markers around the assistant output.

diffusion_gemma_training.jinja

Patched DiffusionGemma template. Diff vs diffusion_gemma.jinja:

Wrap the parts of a model turn that the model emits (thinking channel, tool calls, content, and the closing <turn|>) with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss. Tool responses embedded in the turn come from the environment and stay outside the generation block.

gemma_training.jinja

Patched Gemma template (shared by Gemma and Gemma2, which ship identical chat templates). Diff vs gemma.jinja:

Split the unified assistant output so that the <start_of_turn>model\n header (a prompt cue, not generated by the model) sits outside the generation block, and wrap the assistant content with {% generation %} / {% endgeneration %} markers for SFT assistant-only loss.

gemma3_training.jinja

Patched Gemma 3 template. Same diff as gemma_training.jinja (split the unified output line into role-specific branches so the <start_of_turn>model\n prompt cue sits outside the generation block, and wrap the assistant content with {% generation %} / {% endgeneration %}), applied to the Gemma 3 base template that supports system messages and multimodal content blocks.

glm4moe_training.jinja

Patched GLM-4-MoE template. Diff vs glm4moe.jinja:

Close each assistant turn inside the generation block with the role marker that opens the next turn (<|observation|> after tool calls, otherwise <|user|> or <|system|>). GLM’s original template has no end-of-turn token: a turn is ended by the next message’s role marker, and GLM declares <|user|> and <|observation|> as end-of-sequence tokens, so the training template appends the marker to the assistant turn to keep the stop token in the loss mask, and the following message no longer renders its own marker.

Wrap assistant message output (including the thinking block and tool calls) with {% generation %} / {% endgeneration %} markers for SFT assistant-only loss. With enable_thinking=False, the thinking block stays outside the markers, since the generation prompt already ends with the empty \n<think></think>.

qwen3_training.jinja

Patched Qwen3 template. Diff vs qwen3.jinja:

Always include the thinking block regardless of message position. The original conditionally omits it based on loop.last, which changes the assistant rendering when a tool message is appended, breaking prefix-preservation:

- {%- if loop.index0 > ns.last_query_index %}
-     {%- if loop.last or (not loop.last and reasoning_content) %}
-         {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
-     {%- else %}
-         {{- '<|im_start|>' + message.role + '\n' + content }}
-     {%- endif %}
- {%- else %}
-     {{- '<|im_start|>' + message.role + '\n' + content }}
- {%- endif %}
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}

Wrap assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss. The part of the thinking block that the generation prompt pre-fills (<think>\n\n</think>\n\n with enable_thinking=False) stays outside the markers, since the prompt already ends with it. The other Qwen3 training templates do the same with what their own generation prompt pre-fills, such as <think>\n.

qwen3_v2_training.jinja

Patched original Qwen3 template (shipped by the Qwen3 checkpoints before May 2025, and still by the Qwen3 base checkpoints). Same diff as qwen3_training.jinja, applied to qwen3_v2.jinja.

qwen3_vl_training.jinja

Patched Qwen3-VL template. Diff vs qwen3_vl.jinja:

Wrap assistant message output (both content and tool_calls) with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

gptoss_training.jinja

Patched GPT-OSS template. Diff vs gptoss.jinja:

Wrap assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss. The <|start|>assistant that opens the turn stays outside the markers, since the generation prompt already ends with it.

idefics3_training.jinja

Patched Idefics3 template. Diff vs idefics3.jinja:

Split the assistant message into its own branch so the {% generation %} / {% endgeneration %} markers wrap the assistant content. This enables return_assistant_tokens_mask=True to produce correct masks for SFT assistant-only loss.

lfm2_training.jinja

Patched LFM2 template. Diff vs lfm2.jinja:

Split the unified message output line into role-specific branches, so the <|im_start|>assistant\n prompt cue sits outside the generation block (it is not generated by the model), while the assistant’s content and <|im_end|>\n (which the model must learn to produce and to stop on) sit inside. Wrap the assistant content with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

lfm2_v2_training.jinja

Patched LFM2 template, later revision. Diff vs lfm2_v2.jinja:

Wrap the assistant output with {% generation %} / {% endgeneration %}, in both the branch that renders tool_calls and the plain-content branch, so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss. The <|im_start|>assistant\n prompt cue stays outside the generation block, while the content, the tool calls and <|im_end|>\n sit inside.

llama3_training.jinja

Patched Llama 3 template. Diff vs llama3.jinja:

Wrap assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

llava_next_training.jinja

Patched Llava-Next template. Diff vs llava_next.jinja:

• Wraps assistant message output with {% generation %} / {% endgeneration %} markers for SFT assistant-only loss.
• Keeps the space that follows </s> outside the generation block. The tokenizer merges it into the next turn’s first token (▁[), which would otherwise be masked as assistant output and hide the </s> stop token from the end of the masked span.

mistral_large_4_training.jinja

Patched Mistral Large 4 template. Diff vs mistral_large_4.jinja: same change as mistral_medium_3_5_training.jinja, wrapping the assistant turn (content, tool calls and the end-of-turn token) with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

mistral_medium_3_5_training.jinja

Patched Mistral Medium 3.5 template. Diff vs mistral_medium_3_5.jinja: the original is already prefix-preserving, so the only change is wrapping the assistant turn (content, tool calls and the end-of-turn token) with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

muse_glimmer_training.jinja

Patched Muse Glimmer template. Diff vs muse_glimmer.jinja: wrap the whole assistant branch — every ATEM channel of the turn — with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss. The <|start|>assistant that opens the turn stays outside the markers, since the generation prompt already ends with it.

muse_spark_training.jinja

Patched Muse Spark Open template. Diff vs muse_spark.jinja: the original’s {% generation %} markers also cover the <|start|>assistant that opens the turn, which the generation prompt already ends with. The training template emits it before the markers, so that return_assistant_tokens_mask=True doesn’t train it. The rendered text is unchanged.

nemotron_3_nano_training.jinja

Patched Nemotron Nano template. Diff vs nemotron_3_nano.jinja: the original is already prefix-preserving, so the change is wrapping assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss. The part of the thinking block that the generation prompt pre-fills (<think>\n, or <think></think> with enable_thinking=False) stays outside the markers, since the prompt already ends with it.

nemotron_3_super_training.jinja

Patched Nemotron Super template. Diff vs nemotron_3_super.jinja: same as nemotron_3_nano_training.jinja — the original is already prefix-preserving, so the change is wrapping assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss. The part of the thinking block that the generation prompt pre-fills (<think>\n, or <think></think> with enable_thinking=False) stays outside the markers, since the prompt already ends with it.

nemotron_3_ultra_training.jinja

Patched Nemotron Ultra template. Diff vs nemotron_3_ultra.jinja: same as nemotron_3_nano_training.jinja — the original is already prefix-preserving, so the change is wrapping assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss. The part of the thinking block that the generation prompt pre-fills (<think>\n, or <think></think> with enable_thinking=False) stays outside the markers, since the prompt already ends with it.

nemotron_3_5_lightning_training.jinja

Patched Nemotron 3.5 Lightning template. Diff vs nemotron_3_5_lightning.jinja: same as nemotron_3_nano_training.jinja — the original is already prefix-preserving, so the change is wrapping assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss. The part of the thinking block that the generation prompt pre-fills (<think>\n, or <think></think> with enable_thinking=False) stays outside the markers, since the prompt already ends with it.

phi3_training.jinja

Patched Phi-3 template. Diff vs phi3.jinja:

Wrap assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

phi3_5_training.jinja

Patched Phi-3.5 template. Diff vs phi3_5.jinja:

Wrap assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

qwen2_5_training.jinja

Patched Qwen2.5 template. Diff vs qwen2_5.jinja:

Wrap assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

qwen2_5_base_training.jinja

Patched Qwen2.5 base template (used by the base checkpoints, such as Qwen2.5-0.5B, whose template only differs from the instruct one in its default system prompt). Same diff as qwen2_5_training.jinja, applied to qwen2_5_base.jinja.

qwen2_5_vl_training.jinja

Patched Qwen2.5-VL template (also used for Qwen2-VL, which ships a byte-identical template). Diff vs qwen2_5_vl.jinja:

Split the assistant message into its own branch so the {% generation %} / {% endgeneration %} markers wrap the assistant content. This enables return_assistant_tokens_mask=True to produce correct masks for SFT assistant-only loss.

qwen3_instruct_2507_training.jinja

Patched Qwen3-Instruct-2507 template (used by models like Qwen3-4B-Instruct-2507, which ship a simpler Qwen3 variant without reasoning_content / <think> parsing, multi_step_tool tracking, or the enable_thinking flag). Diff vs qwen3_instruct_2507.jinja:

Wrap assistant message output with {% generation %} / {% endgeneration %} so that return_assistant_tokens_mask=True produces correct masks for SFT assistant-only loss.

qwen3_thinking_2507_training.jinja

Patched Qwen3-Thinking-2507 template (used by models like Qwen3-4B-Thinking-2507, which pre-fill <think>\n in the generation prompt). Diff vs qwen3_thinking_2507.jinja: same set of changes as qwen3_training.jinja, which drops the loop.index0 > ns.last_query_index conditional so the thinking block is always emitted (prefix-preservation) and wraps assistant output with {% generation %} / {% endgeneration %} markers for SFT assistant-only loss.

qwen3_5_think_training.jinja / qwen3_5_nothink_training.jinja

Patched Qwen3.5 templates with shared logic across both flavors (they differ only in the default value of the enable_thinking flag):

• qwen3_5_think_training.jinja defaults to thinking enabled, used by Qwen3.5-4B and larger.
• qwen3_5_nothink_training.jinja defaults to thinking disabled, used by Qwen3.5-2B and smaller.

Diff vs qwen3_5_think.jinja / qwen3_5_nothink.jinja: same as qwen3_training.jinja (drop the loop.index0 > ns.last_query_index conditional so the thinking block is always emitted, wrap assistant output in {% generation %} / {% endgeneration %}).

qwen3_6_training.jinja

Patched Qwen3.6 template. Diff vs qwen3_6.jinja: same as qwen3_training.jinja (drop the loop.index0 > ns.last_query_index conditional so the thinking block is always emitted, wrap assistant output in {% generation %} / {% endgeneration %}).

qwen3_8_training.jinja

Patched Qwen3.8 template. Diff vs qwen3_8.jinja: drop the preserve_thinking / loop.index0 > ns.last_query_index conditional so the thinking block is always emitted — prefix-preservation then holds even when the caller passes preserve_thinking=False — and wrap assistant output with {% generation %} / {% endgeneration %} markers for SFT assistant-only loss, with the part of the thinking block that the generation prompt pre-fills left outside them.

smolvlm_training.jinja

Patched SmolVLM template (also used for SmolVLM2, which ships a byte-identical template). Diff vs smolvlm.jinja:

Split the assistant message into its own branch so the {% generation %} / {% endgeneration %} markers wrap the assistant content. This enables return_assistant_tokens_mask=True to produce correct masks for SFT assistant-only loss.

Related utilities

See Chat Template Utilities for the helper functions (clone_chat_template(), is_chat_template_prefix_preserving, get_training_chat_template()) that operate on these templates.