Add generation markers around answer content for optional SFT loss masking

chat_template.jinja

@@ -1,11 +1,11 @@
-{#- Simple pretraining chat template: question/answer format, all tokens trainable.
-    No chat-ML markers, no thinking tags, no loss masking. -#}
+{#- Pretraining chat template: question/answer format with generation markers.
+    Energon chooses whether to apply SFT loss masking or train on all tokens. -#}
 {%- for message in messages %}
 {%- if message.role == "system" %}
 {{- message.content }}
 {%- elif message.role == "user" %}
 {{- "\nquestion: " + message.content }}
 {%- elif message.role == "assistant" %}
-{{- "\nanswer: " + message.content }}
+{{- "\nanswer: " }}{% generation %}{{- message.content }}{% endgeneration %}
 {%- endif %}
 {%- endfor %}

tokenizer_config.json

@@ -8015,5 +8015,5 @@
   "model_max_length": 262144,
   "tokenizer_class": "PreTrainedTokenizerFast",
   "unk_token": "<unk>",
-  "chat_template": "{#- Simple pretraining chat template: question/answer format, all tokens trainable.\n    No chat-ML markers, no thinking tags, no loss masking. -#}\n{%- for message in messages %}\n{%- if message.role == \"system\" %}\n{{- message.content }}\n{%- elif message.role == \"user\" %}\n{{- \"\\nquestion: \" + message.content }}\n{%- elif message.role == \"assistant\" %}\n{{- \"\\nanswer: \" + message.content }}\n{%- endif %}\n{%- endfor %}\n"
+  "chat_template": "{#- Pretraining chat template: question/answer format with generation markers.\n    Energon chooses whether to apply SFT loss masking or train on all tokens. -#}\n{%- for message in messages %}\n{%- if message.role == \"system\" %}\n{{- message.content }}\n{%- elif message.role == \"user\" %}\n{{- \"\\nquestion: \" + message.content }}\n{%- elif message.role == \"assistant\" %}\n{{- \"\\nanswer: \" }}{% generation %}{{- message.content }}{% endgeneration %}\n{%- endif %}\n{%- endfor %}\n"
 }