Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Fix the ShareGPT mapping for Llama 3.1, Qwen and Gemma chat templates (#12314)

Nilay3ff585ecd112
diff --git a/tests/python/test_get_chat_template_sharegpt_mapping.py b/tests/python/test_get_chat_template_sharegpt_mapping.py
new file mode 100644
index 00000000000..e3f85a5abe6
--- /dev/null
+++ b/tests/python/test_get_chat_template_sharegpt_mapping.py
@@ -0,0 +1,96 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+import pytest
+from tokenizers import Tokenizer, models
+from transformers import PreTrainedTokenizerFast
+
+from unsloth.chat_templates import get_chat_template
+
+
+SHAREGPT = {"role": "from", "content": "value", "user": "human", "assistant": "gpt"}
+SHAREGPT_ROLES = {"system": "system", "user": "human", "assistant": "gpt"}
+CONVERSATION = [
+ ("system", "Be brief."),
+ ("user", "Hi"),
+ ("assistant", "Hello!"),
+ ("user", "Bye"),
+]
+TEMPLATES = [
+ "llama-3.1",
+ "qwen-2.5",
+ "qwen3-instruct",
+ "gemma-3",
+ "gemma-4",
+ "gpt-oss",
+ "llama-3",
+ "chatml",
+ "mistral",
+ "phi-3",
+]
+
+
+def _tokenizer():
+ vocab = {"<unk>": 0, "<s>": 1, "</s>": 2, "<pad>": 3}
+ return PreTrainedTokenizerFast(
+ tokenizer_object = Tokenizer(models.WordLevel(vocab, unk_token = "<unk>")),
+ bos_token = "<s>",
+ eos_token = "</s>",
+ unk_token = "<unk>",
+ pad_token = "<pad>",
+ )
+
+
+def _render(tokenizer, messages, add_generation_prompt):
+ return tokenizer.apply_chat_template(
+ messages, tokenize = False, add_generation_prompt = add_generation_prompt
+ )
+
+
[email protected]("add_generation_prompt", [False, True])
[email protected]("name", TEMPLATES)
+def test_sharegpt_mapping_renders_like_role_content(name, add_generation_prompt):
+ sharegpt = [{"from": SHAREGPT_ROLES[role], "value": text} for role, text in CONVERSATION]
+ role_content = [{"role": role, "content": text} for role, text in CONVERSATION]
+
+ expected = _render(get_chat_template(_tokenizer(), name), role_content, add_generation_prompt)
+ mapped = get_chat_template(_tokenizer(), name, mapping = SHAREGPT)
+
+ assert "Hi" in expected and "Hello!" in expected
+ assert _render(mapped, sharegpt, add_generation_prompt) == expected
+ assert _render(mapped, role_content, add_generation_prompt) == expected
+
+
+def test_sharegpt_mapping_keeps_extra_message_keys():
+ thinking = {"thinking": "Let me think."}
+ conversation = CONVERSATION[:3]
+ sharegpt = [
+ dict(
+ {"from": SHAREGPT_ROLES[role], "value": text},
+ **(thinking if role == "assistant" else {}),
+ )
+ for role, text in conversation
+ ]
+ role_content = [
+ dict({"role": role, "content": text}, **(thinking if role == "assistant" else {}))
+ for role, text in conversation
+ ]
+
+ expected = _render(get_chat_template(_tokenizer(), "gpt-oss"), role_content, False)
+ mapped = get_chat_template(_tokenizer(), "gpt-oss", mapping = SHAREGPT)
+
+ assert "Let me think." in expected
+ assert _render(mapped, sharegpt, False) == expected
+
+
+def test_custom_template_reading_sharegpt_keys_keeps_working():
+ template = (
+ "{{ bos_token }}{% for message in messages %}"
+ "{% if message['from'] == 'human' %}{{ 'User: ' + message['value'] + '\n' }}"
+ "{% elif message['from'] == 'gpt' %}{{ 'Bot: ' + message['value'] + eos_token + '\n' }}"
+ "{% endif %}{% endfor %}"
+ )
+ sharegpt = [{"from": SHAREGPT_ROLES[role], "value": text} for role, text in CONVERSATION]
+ mapped = get_chat_template(_tokenizer(), (template, "eos_token"), mapping = SHAREGPT)
+
+ assert _render(mapped, sharegpt, False) == "<s>User: Hi\nBot: Hello!</s>\nUser: Bye\n"
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index 7f9933ed96b..c297f89237c 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -2094,13 +2094,6 @@ def get_chat_template(
if IS_GEMMA and not chat_template.startswith(("{{ bos_token }}", "{{- bos_token }}")):
chat_template = "{{ bos_token }}" + chat_template
- # The spliced ShareGPT values land inside Jinja literals, so escape them.
- new_chat_template = chat_template\
- .replace("'role'", "'" + _escape_jinja_literal(mapping["role"]) + "'")\
- .replace("'content'", "'" + _escape_jinja_literal(mapping["content"]) + "'")\
- .replace("'user'", "'" + _escape_jinja_literal(mapping["user"]) + "'")\
- .replace("'assistant'", "'" + _escape_jinja_literal(mapping["assistant"]) + "'")
-
if use_zoo_tokenizer_patch:
# Unsloth MLX avoids the model-utils tokenizer wrapper: that import path pulls Torch/GPU-specific
# modules in before MLX training.
@@ -2112,14 +2105,22 @@ def get_chat_template(
# If not normal HF, we add a check to make old templates work
if mapping != {"role" : "role", "content" : "content", "user" : "user", "assistant" : "assistant"}:
+ role, content, user, assistant = (
+ "'" + _escape_jinja_literal(mapping[key]) + "'"
+ for key in ("role", "content", "user", "assistant")
+ )
chat_template = \
- "{% if 'role' in messages[0] %}" + \
- chat_template + \
- "{% else %}" + \
- new_chat_template + \
- "{% endif %}"
- else:
- chat_template = new_chat_template
+ "{%- if 'role' not in messages[0] -%}" + \
+ "{%- set sharegpt = namespace(messages = []) -%}" + \
+ "{%- for message in messages -%}" + \
+ "{%- set role = {" + user + " : 'user', " + assistant + " : 'assistant'}" + \
+ ".get(message[" + role + "], message[" + role + "]) -%}" + \
+ "{%- set sharegpt.messages = sharegpt.messages + " + \
+ "[dict(message, role = role, content = message[" + content + "])] -%}" + \
+ "{%- endfor -%}" + \
+ "{%- set messages = sharegpt.messages -%}" + \
+ "{%- endif %}" + \
+ chat_template
chat_template, system_message = _change_system_message(chat_template, type_chat_template, system_message)