Update chat_template.jinja
Browse files- chat_template.jinja +59 -35
chat_template.jinja
CHANGED
@@ -117,14 +117,15 @@
|
|
117 |
{%- set tool = tool.function %}
|
118 |
{{- "// " + tool.description + "
|
119 |
" }}
|
120 |
-
{{- "type "+ tool.name + " =
|
121 |
-
{%- if tool.parameters and tool.parameters.properties
|
122 |
-
{{- "_:
|
123 |
-
{{- "{
|
124 |
" }}
|
125 |
{%- for param_name, param_spec in tool.parameters.properties.items() %}
|
126 |
-
{
|
|
|
127 |
" }}
|
|
|
128 |
{{- param_name }}
|
129 |
{%- if param_name not in (tool.parameters.required or []) -%}
|
130 |
{{- "?" }}
|
@@ -132,7 +133,9 @@
|
|
132 |
{{- ": " }}
|
133 |
{{- render_typescript_type(param_spec, tool.parameters.required or []) }}
|
134 |
{%- if param_spec.default is defined -%}
|
135 |
-
{%- if param_spec.
|
|
|
|
|
136 |
{{- "// default: " + param_spec.default }}
|
137 |
{%- else %}
|
138 |
{{- ", // default: " + param_spec.default|tojson }}
|
@@ -140,20 +143,22 @@
|
|
140 |
{%- endif -%}
|
141 |
{%- if not loop.last %}
|
142 |
{{- ",
|
|
|
|
|
|
|
143 |
" }}
|
144 |
{%- endif -%}
|
145 |
{%- endfor %}
|
146 |
-
{{- "
|
147 |
-
|
148 |
" }}
|
149 |
{%- else -%}
|
150 |
-
{{- "
|
151 |
-
|
152 |
" }}
|
153 |
{%- endif -%}
|
154 |
{%- endfor %}
|
155 |
-
{{- "
|
156 |
-
} // namespace " + namespace_name }}
|
157 |
{%- endmacro -%}
|
158 |
|
159 |
{%- macro render_builtin_tools(browser_tool, python_tool) -%}
|
@@ -250,11 +255,10 @@
|
|
250 |
{#- System Message Construction ============================================ #}
|
251 |
{%- macro build_system_message() -%}
|
252 |
{%- if model_identity is not defined %}
|
253 |
-
{
|
254 |
-
" -}}
|
255 |
-
{%- else %}
|
256 |
-
{{- model_identity }}
|
257 |
{%- endif %}
|
|
|
|
|
258 |
{{- "Knowledge cutoff: 2024-06
|
259 |
" }}
|
260 |
{{- "Current date: " + strftime_now("%Y-%m-%d") + "
|
@@ -263,7 +267,7 @@
|
|
263 |
{%- if reasoning_effort is not defined %}
|
264 |
{%- set reasoning_effort = "medium" %}
|
265 |
{%- endif %}
|
266 |
-
{{- "
|
267 |
|
268 |
" }}
|
269 |
{%- if builtin_tools %}
|
@@ -280,9 +284,11 @@
|
|
280 |
{%- endfor %}
|
281 |
{{- render_builtin_tools(available_builtin_tools.browser, available_builtin_tools.python) }}
|
282 |
{%- endif -%}
|
283 |
-
{{- "# Valid channels: analysis, commentary, final. Channel must be included for every message.
|
284 |
-
|
285 |
-
|
|
|
|
|
286 |
{%- endmacro -%}
|
287 |
|
288 |
{#- Main Template Logic ================================================= #}
|
@@ -328,42 +334,60 @@
|
|
328 |
{%- for message in loop_messages -%}
|
329 |
{#- At this point only assistant/user/tool messages should remain #}
|
330 |
{%- if message.role == 'assistant' -%}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
331 |
{%- if "tool_calls" in message %}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
332 |
{#- We assume max 1 tool call per message, and so we infer the tool call name #}
|
333 |
{#- in "tool" messages from the most recent assistant tool call name #}
|
334 |
{%- set tool_call = message.tool_calls[0] %}
|
335 |
{%- if tool_call.function %}
|
336 |
{%- set tool_call = tool_call.function %}
|
337 |
{%- endif %}
|
338 |
-
{%- if message.content %}
|
|
|
|
|
339 |
{{- "<|start|>assistant<|channel|>analysis<|message|>" + message.content + "<|end|>" }}
|
|
|
|
|
340 |
{%- endif %}
|
341 |
{{- "<|start|>assistant to=" }}
|
342 |
-
{{- "functions." + tool_call.name + "<|channel|>commentary
|
|
|
343 |
{{- tool_call.arguments|tojson }}
|
344 |
-
{{- "<|
|
345 |
{%- set last_tool_call.name = tool_call.name %}
|
346 |
-
{%- elif
|
347 |
{#- Only render the CoT if the final turn is an assistant turn and add_generation_prompt is false #}
|
348 |
{#- This is a situation that should only occur in training, never in inference. #}
|
349 |
-
{
|
|
|
|
|
350 |
{#- <|return|> indicates the end of generation, but <|end|> does not #}
|
351 |
{#- <|return|> should never be an input to the model, but we include it as the final token #}
|
352 |
{#- when training, so the model learns to emit it. #}
|
353 |
{{- "<|start|>assistant<|channel|>final<|message|>" + message.content + "<|return|>" }}
|
354 |
-
|
355 |
-
{%- elif "thinking" in message %}
|
356 |
{#- CoT is dropped during all previous turns, so we never render it for inference #}
|
357 |
{{- "<|start|>assistant<|channel|>final<|message|>" + message.content + "<|end|>" }}
|
358 |
{%- set last_tool_call.name = none %}
|
359 |
-
{%- elif loop.last and not add_generation_prompt %}
|
360 |
-
{#- <|return|> indicates the end of generation, but <|end|> does not #}
|
361 |
-
{#- <|return|> should never be an input to the model, but we include it as the final token #}
|
362 |
-
{#- when training, so the model learns to emit it. #}
|
363 |
-
{{- "<|start|>assistant<|message|>" + message.content + "<|return|>" }}
|
364 |
-
{%- else %}
|
365 |
-
{{- "<|start|>assistant<|message|>" + message.content + "<|end|>" }}
|
366 |
-
{%- set last_tool_call.name = none %}
|
367 |
{%- endif %}
|
368 |
{%- elif message.role == 'tool' -%}
|
369 |
{%- if last_tool_call.name is none %}
|
|
|
117 |
{%- set tool = tool.function %}
|
118 |
{{- "// " + tool.description + "
|
119 |
" }}
|
120 |
+
{{- "type "+ tool.name + " = " }}
|
121 |
+
{%- if tool.parameters and tool.parameters.properties %}
|
122 |
+
{{- "(_: {
|
|
|
123 |
" }}
|
124 |
{%- for param_name, param_spec in tool.parameters.properties.items() %}
|
125 |
+
{%- if param_spec.description %}
|
126 |
+
{{- "// " + param_spec.description + "
|
127 |
" }}
|
128 |
+
{%- endif %}
|
129 |
{{- param_name }}
|
130 |
{%- if param_name not in (tool.parameters.required or []) -%}
|
131 |
{{- "?" }}
|
|
|
133 |
{{- ": " }}
|
134 |
{{- render_typescript_type(param_spec, tool.parameters.required or []) }}
|
135 |
{%- if param_spec.default is defined -%}
|
136 |
+
{%- if param_spec.enum %}
|
137 |
+
{{- ", // default: " + param_spec.default }}
|
138 |
+
{%- elif param_spec.oneOf %}
|
139 |
{{- "// default: " + param_spec.default }}
|
140 |
{%- else %}
|
141 |
{{- ", // default: " + param_spec.default|tojson }}
|
|
|
143 |
{%- endif -%}
|
144 |
{%- if not loop.last %}
|
145 |
{{- ",
|
146 |
+
" }}
|
147 |
+
{%- else %}
|
148 |
+
{{- ",
|
149 |
" }}
|
150 |
{%- endif -%}
|
151 |
{%- endfor %}
|
152 |
+
{{- "}) => any;
|
153 |
+
|
154 |
" }}
|
155 |
{%- else -%}
|
156 |
+
{{- "() => any;
|
157 |
+
|
158 |
" }}
|
159 |
{%- endif -%}
|
160 |
{%- endfor %}
|
161 |
+
{{- "} // namespace " + namespace_name }}
|
|
|
162 |
{%- endmacro -%}
|
163 |
|
164 |
{%- macro render_builtin_tools(browser_tool, python_tool) -%}
|
|
|
255 |
{#- System Message Construction ============================================ #}
|
256 |
{%- macro build_system_message() -%}
|
257 |
{%- if model_identity is not defined %}
|
258 |
+
{%- set model_identity = "You are ChatGPT, a large language model trained by OpenAI." %}
|
|
|
|
|
|
|
259 |
{%- endif %}
|
260 |
+
{{- model_identity + "
|
261 |
+
" }}
|
262 |
{{- "Knowledge cutoff: 2024-06
|
263 |
" }}
|
264 |
{{- "Current date: " + strftime_now("%Y-%m-%d") + "
|
|
|
267 |
{%- if reasoning_effort is not defined %}
|
268 |
{%- set reasoning_effort = "medium" %}
|
269 |
{%- endif %}
|
270 |
+
{{- "Reasoning: " + reasoning_effort + "
|
271 |
|
272 |
" }}
|
273 |
{%- if builtin_tools %}
|
|
|
284 |
{%- endfor %}
|
285 |
{{- render_builtin_tools(available_builtin_tools.browser, available_builtin_tools.python) }}
|
286 |
{%- endif -%}
|
287 |
+
{{- "# Valid channels: analysis, commentary, final. Channel must be included for every message." }}
|
288 |
+
{%- if tools -%}
|
289 |
+
{{- "
|
290 |
+
Calls to these tools must go to the commentary channel: 'functions'." }}
|
291 |
+
{%- endif -%}
|
292 |
{%- endmacro -%}
|
293 |
|
294 |
{#- Main Template Logic ================================================= #}
|
|
|
334 |
{%- for message in loop_messages -%}
|
335 |
{#- At this point only assistant/user/tool messages should remain #}
|
336 |
{%- if message.role == 'assistant' -%}
|
337 |
+
{#- Checks to ensure the messages are being passed in the format we expect #}
|
338 |
+
{%- if "content" in message %}
|
339 |
+
{%- if "<|channel|>analysis<|message|>" in message.content or "<|channel|>final<|message|>" in message.content %}
|
340 |
+
{{- raise_exception("You have passed a message containing <|channel|> tags in the content field. Instead of doing this, you should pass analysis messages (the string between '<|message|>' and '<|end|>') in the 'thinking' field, and final messages (the string between '<|message|>' and '<|end|>') in the 'content' field.") }}
|
341 |
+
{%- endif %}
|
342 |
+
{%- endif %}
|
343 |
+
{%- if "thinking" in message %}
|
344 |
+
{%- if "<|channel|>analysis<|message|>" in message.thinking or "<|channel|>final<|message|>" in message.thinking %}
|
345 |
+
{{- raise_exception("You have passed a message containing <|channel|> tags in the thinking field. Instead of doing this, you should pass analysis messages (the string between '<|message|>' and '<|end|>') in the 'thinking' field, and final messages (the string between '<|message|>' and '<|end|>') in the 'content' field.") }}
|
346 |
+
{%- endif %}
|
347 |
+
{%- endif %}
|
348 |
{%- if "tool_calls" in message %}
|
349 |
+
{#- We need very careful handling here - we want to drop the tool call analysis message if the model #}
|
350 |
+
{#- has output a later <|final|> message, but otherwise we want to retain it. This is the only case #}
|
351 |
+
{#- when we render CoT/analysis messages in inference. #}
|
352 |
+
{%- set future_final_message = namespace(found=false) %}
|
353 |
+
{%- for future_message in loop_messages[loop.index:] %}
|
354 |
+
{%- if future_message.role == 'assistant' and "tool_calls" not in future_message %}
|
355 |
+
{%- set future_final_message.found = true %}
|
356 |
+
{%- endif %}
|
357 |
+
{%- endfor %}
|
358 |
{#- We assume max 1 tool call per message, and so we infer the tool call name #}
|
359 |
{#- in "tool" messages from the most recent assistant tool call name #}
|
360 |
{%- set tool_call = message.tool_calls[0] %}
|
361 |
{%- if tool_call.function %}
|
362 |
{%- set tool_call = tool_call.function %}
|
363 |
{%- endif %}
|
364 |
+
{%- if message.content and message.thinking %}
|
365 |
+
{{- raise_exception("Cannot pass both content and thinking in an assistant message with tool calls! Put the analysis message in one or the other, but not both.") }}
|
366 |
+
{%- elif message.content and not future_final_message.found %}
|
367 |
{{- "<|start|>assistant<|channel|>analysis<|message|>" + message.content + "<|end|>" }}
|
368 |
+
{%- elif message.thinking and not future_final_message.found %}
|
369 |
+
{{- "<|start|>assistant<|channel|>analysis<|message|>" + message.thinking + "<|end|>" }}
|
370 |
{%- endif %}
|
371 |
{{- "<|start|>assistant to=" }}
|
372 |
+
{{- "functions." + tool_call.name + "<|channel|>commentary " }}
|
373 |
+
{{- (tool_call.content_type if tool_call.content_type is defined else "json") + "<|message|>" }}
|
374 |
{{- tool_call.arguments|tojson }}
|
375 |
+
{{- "<|call|>" }}
|
376 |
{%- set last_tool_call.name = tool_call.name %}
|
377 |
+
{%- elif loop.last and not add_generation_prompt %}
|
378 |
{#- Only render the CoT if the final turn is an assistant turn and add_generation_prompt is false #}
|
379 |
{#- This is a situation that should only occur in training, never in inference. #}
|
380 |
+
{%- if "thinking" in message %}
|
381 |
+
{{- "<|start|>assistant<|channel|>analysis<|message|>" + message.thinking + "<|end|>" }}
|
382 |
+
{%- endif %}
|
383 |
{#- <|return|> indicates the end of generation, but <|end|> does not #}
|
384 |
{#- <|return|> should never be an input to the model, but we include it as the final token #}
|
385 |
{#- when training, so the model learns to emit it. #}
|
386 |
{{- "<|start|>assistant<|channel|>final<|message|>" + message.content + "<|return|>" }}
|
387 |
+
{%- else %}
|
|
|
388 |
{#- CoT is dropped during all previous turns, so we never render it for inference #}
|
389 |
{{- "<|start|>assistant<|channel|>final<|message|>" + message.content + "<|end|>" }}
|
390 |
{%- set last_tool_call.name = none %}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
391 |
{%- endif %}
|
392 |
{%- elif message.role == 'tool' -%}
|
393 |
{%- if last_tool_call.name is none %}
|