Differences

This shows you the differences between two versions of the page.

--- 712:configure_stages [2026/02/26 09:48] – Prinz, Patrick
+++ 712:configure_stages [2026/03/25 09:44] (current) – [AI Feature Tokenizer Configuration] Prinz, Patrick
@@ Line 526: / Line 526: @@
 </cg>
 </code>
+**When using models newer than gpt4o (like gpt5)**
+>The temperature has to be adjusted to the default value of 1. Add <cg-property name="temperature" value="1" /> to the relevant <cg-hosts>
+>The <chatbot-property name="maxTokens" value="500"></chatbot-property> has to be replaced with <chatbot-property name="maxCompletionTokens" value="500"></chatbot-property>
 **AI content generation feature (OpenAI compatible):**
@@ Line 683: / Line 689: @@
   <cg-property name="proxy" value="<proxy_ident>" />
 <cg-host/>
+</code>
+===== AI Feature Tokenizer Configuration =====
+LLM that are NOT using the standard byte-pair-encoding algorithm or are provided via the openAICompatible adapter CANNOT be used with the Azure tokenizer. Stages implements the HuggingFaceTokenizer from deep java library to support customized tokenizer.
+Add the property “tokenizerFolder” to the cg-host section
+   <cg-host ident="custom-model-ident" url="${ai.model.url}" displayName="custom-model">
+    <cg-property name=".... />
+    <cg-property name="tokenizerFolder" value="custom-model-v1"/>
+  </cg-host>
+ In methodpark\stages\conf
+Create a folder tokenizer\custom-model-v1 ← make sure the folder name exactly matches the value in the tokenizerFolder property
+Add tokenizer.json and tokenizer_config.json
+**Example for using DeepSeek V3.2**
+. cg-host configuration in config.xml
+<code ->
+<cg-type name="openAICompatible">
+ <cg-host ident="training-model" url="${ai.model.url}" displayName="DeepSeek V3.2 for Chat">
+  <cg-property name="deployment_name" value="DeepSeek_V3_2"/>
+  <cg-property name="tokenizerFolder" value="DeepSeek_V3_2"/>
+  ...
+</cg-host>
+<cg-type />
+</code>
+. chatbot configuration in config.xml
+>You need to change all chatbot steps that shall use the configured LLM
+<code ->
+default:
+<chatbot-step ident="summaries" type="azureOpenAI" host="gpt-4o-mini">
+custom configuration:
+<chatbot-step ident="summaries" type="openAICompatible" host="training-model">
+</code>
+. Add tokenizer configuration:
+As the custom models are not processed by the standard azure tokenizer, there has to be a custom configuration added and put into a folder in conf/tokenizer that matches the config.xml (e.g. conf/tokenizer/DeepSeek_V3_2). The custom tokenizer can use tokenizer.json and tokenizer_config.json files.
+For lots of models various configurations can be downloaded from Huggingface.co, e.g.:
+  * [[https://huggingface.co/deepseek-ai/DeepSeek-V3.2/blob/main/tokenizer.json]]
+  * [[https://huggingface.co/deepseek-ai/DeepSeek-V3.2/blob/main/tokenizer_config.json]]
+To make the tokenizer_config.json work with Stages a few small adjustments have to be made
+>In tokenizer_config.json look up objects (e.g. bos_token) and just use the value from the “content” attribute value instead of the whole object. See the following file examples:
+**Original tokenizer_config.json :**
+<code ->
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "bos_token": {
+    "__type": "AddedToken",
+    "content": "<｜begin▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "clean_up_tokenization_spaces": false,
+  "eos_token": {
+    "__type": "AddedToken",
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "legacy": true,
+  "model_max_length": 16384,
+  "pad_token": {
+    "__type": "AddedToken",
+    "content": "<｜end▁of▁sentence｜>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "sp_model_kwargs": {},
+  "unk_token": null,
+  "tokenizer_class": "LlamaTokenizerFast",
+  "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='', is_first_sp=true) %}{%- for message in messages %}{%- if message['role'] == 'system' %}{%- if ns.is_first_sp %}{% set ns.system_prompt = ns.system_prompt + message['content'] %}{% set ns.is_first_sp = false %}{%- else %}{% set ns.system_prompt = ns.system_prompt + '\\n\\n' + message['content'] %}{%- endif %}{%- endif %}{%- endfor %}{{ bos_token }}{{ ns.system_prompt }}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' in message %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls'] %}{%- if not ns.is_first %}{%- if message['content'] is none %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- else %}{{'<｜Assistant｜>' + message['content'] + '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- set ns.is_first = true -%}{%- else %}{{'\\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- endfor %}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' not in message %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜><think>\\n'}}{% endif %}"
+}
+</code>
+**Updated tokenizer_config.json**
+<code ->
+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "bos_token": "<｜begin▁of▁sentence｜>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<｜end▁of▁sentence｜>",
+  "legacy": true,
+  "model_max_length": 16384,
+  "pad_token": "<｜end▁of▁sentence｜>",
+  "unk_token": null,
+  "tokenizer_class": "LlamaTokenizerFast",
+  "chat_template": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='', is_first_sp=true) %}{%- for message in messages %}{%- if message['role'] == 'system' %}{%- if ns.is_first_sp %}{% set ns.system_prompt = ns.system_prompt + message['content'] %}{% set ns.is_first_sp = false %}{%- else %}{% set ns.system_prompt = ns.system_prompt + '\\n\\n' + message['content'] %}{%- endif %}{%- endif %}{%- endfor %}{{ bos_token }}{{ ns.system_prompt }}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' in message %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls'] %}{%- if not ns.is_first %}{%- if message['content'] is none %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- else %}{{'<｜Assistant｜>' + message['content'] + '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- set ns.is_first = true -%}{%- else %}{{'\\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- endfor %}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' not in message %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜><think>\\n'}}{% endif %}"
+}
 </code>