Trying to get to the bottom of why `Qwen1.5-110B-Chat` scores so much higher than the `command-r` models

[jukofyork]

https://github.com/EQ-bench/EQ-Bench/blob/main_v2_4/instruction-templates/Cohere.yaml

user: "<|USER_TOKEN|>"
bot: "<|CHATBOT_TOKEN|>"
turn_template: "<|START_OF_TURN_TOKEN|><|user|><|user-message|><|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|bot|><|bot-message|>"
context: "<BOS_TOKEN>"
system_message: ""

Is there a chance that the "<BOS_TOKEN> was getting added twice during the tests:

https://huggingface.co/CohereForAI/c4ai-command-r-plus/discussions/22#66179da37ed574892089967c

Not sure which backend was used to test against, but some will add is automatically and the HF config looks to have it both:

  "add_bos_token": true

and:

"chat_template": [
    {
      "name": "default",
      "template": "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% elif false == true %}{% set loop_messages = messages %}{% set system_message = 'You are Command-R, a brilliant, sophisticated, AI-assistant trained to assist human users by providing thorough responses. You are trained by Cohere.' %}{% else %}{% set loop_messages = messages %}{% set system_message = false %}{% endif %}{% if system_message != false %}{{ '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' + system_message + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% for message in loop_messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|START_OF_TURN_TOKEN|><|USER_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% elif message['role'] == 'assistant' %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>'  + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}{% endif %}"
    },
    {
      "name": "tool_use",
      "template": "\n{%- macro json_to_python_type(json_spec) %}\n{%- set basic_type_map = {\n    \"string\": \"str\",\n    \"number\": \"float\",\n    \"integer\": \"int\",\n    \"boolean\": \"bool\"\n} %}\n\n{%- if basic_type_map[json_spec.type] is defined %}\n    {{- basic_type_map[json_spec.type] }}\n{%- elif json_spec.type == \"array\" %}\n    {{- \"List[\" +  json_to_python_type(json_spec.items) + \"]\"}}\n{%- elif json_spec.type == \"object\" %}\n    {{- \"Dict[str, \" + json_to_python_type(json_spec.additionalProperties) + ']'}}\n{%- elif json_spec.type is iterable %}\n    {{- \"Union[\" }}\n    {%- for t in json_spec.type %}\n      {{- json_to_python_type({\"type\": t}) }}\n      {%- if not loop.last %}\n        {{- \",\" }} \n    {%- endif %}\n    {%- endfor %}\n    {{- \"]\" }}\n{%- else %}\n    {{- \"Any\" }}\n{%- endif %}\n{%- endmacro %}\n\n{%- macro old_tool_parser(tools) %}\n{%- for tool in tools %}\n    {%- if loop.index0 != 0 %}\n        {{- '\\n\\n' }}\n    {%- endif %}\n    {{- '```python\\ndef ' + tool.name + '(' }}\n    {%- for param_name, param_fields in tool.parameter_definitions|items %}\n        {%- if loop.index0 != 0 %}\n            {{- ', '}}\n        {%- endif %}\n        {{- param_name + ': ' }}\n        {%- if not param_fields.required %}\n            {{- 'Optional[' + param_fields.type + '] = None'}}\n        {%- else %}\n            {{- param_fields.type }}\n        {%- endif %}\n    {%- endfor %}\n    {{- ') -> List[Dict]:\\n    \"\"\"'}}\n    {{- tool.description }}\n    {%- if tool.parameter_definitions|length != 0 %}\n        {{- '\\n\\n    Args:\\n        '}}\n        {%- for param_name, param_fields in tool.parameter_definitions|items %}\n            {%- if loop.index0 != 0 %}\n                {{- '\\n        ' }}\n            {%- endif %}\n            {{- param_name + ' ('}}\n            {%- if not param_fields.required %}\n                {{- 'Optional[' + param_fields.type + ']'}}\n            {%- else %}\n                {{- param_fields.type }}\n            {%- endif %}\n            {{- '): ' + param_fields.description }}\n        {%- endfor %}\n    {%- endif %}\n    {{- '\\n    \"\"\"\\n    pass\\n```' }}\n{%- endfor %}\n{%- endmacro %}\n\n{%- macro new_tool_parser(tools) %}\n{%- for tool in tools %}\n  {%- if loop.index0 != 0 %}\n    {{- '\\n\\n'}}\n  {%- endif %}\n  {%- if tool.function is defined %}\n    {%- set tool = tool.function %}\n  {%- endif %}\n  {{-'```python\ndef ' + tool.name + '('}}\n  {%- for param_name, param_fields in tool.parameters.properties|items %}\n    {%- if loop.index0 != 0 %}\n      {{- ', '}}\n    {%- endif %}\n    {{-param_name + \": \"}} \n    {%- if not param_name in tool.parameters.required %}\n      {{-'Optional[' + json_to_python_type(param_fields) + '] = None'}}\n    {%- else %}\n      {{- json_to_python_type(param_fields) }}\n    {%- endif %}\n  {%- endfor %}\n  {{- ') -> List[Dict]:\n    \"\"\"'}}\n  {{- tool.description }}\n  {%- if tool.parameters.properties|length != 0 %}\n    {{- '\\n\\n    Args:\\n        '}}\n    {%- for param_name, param_fields in tool.parameters.properties|items %}\n      {%- if loop.index0 != 0 %}\n        {{- '\\n        ' }}\n      {%- endif %}\n      {{- param_name + ' ('}}\n      {%- if not param_name in tool.parameters.required %}\n        {{-'Optional[' + json_to_python_type(param_fields) + ']'}}\n      {%- else %}\n        {{- json_to_python_type(param_fields) }}\n      {%- endif %}\n      {{- '): ' + param_fields.description }}\n    {%- endfor %}\n    {%- endif %}\n    {{- '\\n    \"\"\"\\n    pass\\n```' }}\n{%- endfor %}\n{%- endmacro %}\n\n{{- bos_token }}\n{%- if messages[0]['role'] == 'system' %}\n  {%- set loop_messages = messages[1:] %}\n  {%- set system_message = messages[0]['content'] %}\n{%- else %}\n  {%- set loop_messages = messages %}\n  {%- set system_message = '## Task and Context\\nYou help people answer their questions and other requests interactively. You will be asked a very wide array of requests on all kinds of topics. You will be equipped with a wide range of search engines or similar tools to help you, which you use to research your answer. You should focus on serving the user\\'s needs as best you can, which will be wide-ranging.\\n\\n## Style Guide\\nUnless the user asks for a different style of answer, you should answer in full sentences, using proper grammar and spelling.' %}\n{%- endif %}\n{{- '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' }}\n{{- '# Safety Preamble' }}\n{{- '\nThe instructions in this section override those in the task description and style guide sections. Don\\'t answer questions that are harmful or immoral.' }}\n{{- '\n\n# System Preamble' }}\n{{- '\n## Basic Rules' }}\n{{- '\nYou are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user\\'s requests, you cite your sources in your answers, according to those instructions.' }}\n{{- '\n\n# User Preamble' }}\n{{- '\n' + system_message }}\n{{-'\n\n## Available Tools\nHere is a list of tools that you have available to you:\n\n'}}\n{%- set ns = namespace(new_tools=true) %}\n{%- for tool in tools %}\n    {%- if tool.parameter_definitions is defined %}\n        {%- set ns.new_tools = false %}\n    {%- endif %}\n{%- endfor %}\n{%- if ns.new_tools %}\n    {{- new_tool_parser(tools) }}\n{%- else %}\n    {{- old_tool_parser(tools) }}\n{%- endif %}\n{{- '<|END_OF_TURN_TOKEN|>'}}\n{%- for message in loop_messages %}\n  {%- set content = message['content'] %}\n  {%- if message.role == 'user' %}\n    {{- '<|START_OF_TURN_TOKEN|><|USER_TOKEN|>' + content|trim + '<|END_OF_TURN_TOKEN|>' }}\n  {%- elif message.role == 'system' %}\n    {{- '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' + content|trim + '<|END_OF_TURN_TOKEN|>' }}\n  {%- elif message.role == 'assistant' and message.tool_calls is defined %}\n    {{- '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}\n    {%- if message.content is defined %}\n        {{- message.content|trim }}\n    {%- endif %}\n    {{- '\\nAction:\\n```json\\n[\\n' }}\n    {%- for tool_call in message.tool_calls %}\n        {%- if tool_call.function is defined %}\n            {%- set tool_call = tool_call.function %}\n        {%- endif %}\n        {{- '{\\n'|indent(4, first=true) }}\n        {{- '\"tool_name\": \"'|indent(8, first=true) + tool_call.name + '\",\\n' }}\n        {{- '\"parameters\": '|indent(8, first=true) }}\n        {%- if tool_call.arguments is defined and tool_call.arguments|length > 0 %}    \n            {{- tool_call.arguments|tojson(indent=4)|indent(8) }}\n            {{- '\\n' }}\n        {%- else %}\n            {{- '{}\\n' }}\n        {%- endif %}\n        {{- '}'|indent(4, first=true) }}\n        {%- if not loop.last %}\n            {{- ',\\n' }}\n        {%- endif %}\n    {%- endfor %}\n    {{- \"\\n]```\\n\" }}\n  {%- elif message.role == 'assistant' %}\n    {{- '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>'  + content|trim + '<|END_OF_TURN_TOKEN|>' }}\n  {%- elif message.role == 'tool' %}\n    {{- '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|><results>\\n' }}\n    {{- message.content|trim }}\n    {{- '</results><|END_OF_TURN_TOKEN|>' }}\n  {%- endif %}\n{%- endfor %}\n{{-'<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>Write \\'Action:\\' followed by a json-formatted list of actions that you want to perform in order to produce a good response to the user\\'s last input. You can use any of the supplied tools any number of times, but you should aim to execute the minimum number of necessary actions for the input. You should use the `directly-answer` tool if calling the other tools is unnecessary. The list of actions you want to call should be formatted as a list of json objects, for example:\n```json\n[\n    {\n        \"tool_name\": title of the tool in the specification,\n        \"parameters\": a dict of parameters to input into the tool as they are defined in the specs, or {} if it takes no parameters\n    }\n]```<|END_OF_TURN_TOKEN|>'}}\n{%- if add_generation_prompt %}\n  {{- '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}\n{%- endif %}\n"
    },
    {
      "name": "rag",
      "template": "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% set system_message = '## Task and Context\\nYou help people answer their questions and other requests interactively. You will be asked a very wide array of requests on all kinds of topics. You will be equipped with a wide range of search engines or similar tools to help you, which you use to research your answer. You should focus on serving the user\\'s needs as best you can, which will be wide-ranging.\\n\\n## Style Guide\\nUnless the user asks for a different style of answer, you should answer in full sentences, using proper grammar and spelling.' %}{% endif %}{{ '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' }}{{ '# Safety Preamble' }}{{ '\nThe instructions in this section override those in the task description and style guide sections. Don\\'t answer questions that are harmful or immoral.' }}{{ '\n\n# System Preamble' }}{{ '\n## Basic Rules' }}{{ '\nYou are a powerful conversational AI trained by Cohere to help people. You are augmented by a number of tools, and your job is to use and consume the output of these tools to best help the user. You will see a conversation history between yourself and a user, ending with an utterance from the user. You will then see a specific instruction instructing you what kind of response to generate. When you answer the user\\'s requests, you cite your sources in your answers, according to those instructions.' }}{{ '\n\n# User Preamble' }}{{ '\n' + system_message }}{{ '<|END_OF_TURN_TOKEN|>'}}{% for message in loop_messages %}{% set content = message['content'] %}{% if message['role'] == 'user' %}{{ '<|START_OF_TURN_TOKEN|><|USER_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% elif message['role'] == 'system' %}{{ '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% elif message['role'] == 'assistant' %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>'  + content.strip() + '<|END_OF_TURN_TOKEN|>' }}{% endif %}{% endfor %}{{ '<|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>'}}{{ '<results>' }}{% for document in documents %}{{ '\nDocument: ' }}{{ loop.index0 }}\n{% for key, value in document.items() %}{{ key }}: {{value}}\n{% endfor %}{% endfor %}{{ '</results>'}}{{ '<|END_OF_TURN_TOKEN|><|START_OF_TURN_TOKEN|><|SYSTEM_TOKEN|>' }}{{ 'Carefully perform the following instructions, in order, starting each with a new line.\n' }}{{ 'Firstly, Decide which of the retrieved documents are relevant to the user\\'s last input by writing \\'Relevant Documents:\\' followed by comma-separated list of document numbers. If none are relevant, you should instead write \\'None\\'.\n' }}{{ 'Secondly, Decide which of the retrieved documents contain facts that should be cited in a good answer to the user\\'s last input by writing \\'Cited Documents:\\' followed a comma-separated list of document numbers. If you dont want to cite any of them, you should instead write \\'None\\'.\n' }}{% if citation_mode=='accurate' %}{{ 'Thirdly, Write \\'Answer:\\' followed by a response to the user\\'s last input in high quality natural english. Use the retrieved documents to help you. Do not insert any citations or grounding markup.\n' }}{% endif %}{{ 'Finally, Write \\'Grounded answer:\\' followed by a response to the user\\'s last input in high quality natural english. Use the symbols <co: doc> and </co: doc> to indicate when a fact comes from a document in the search result, e.g <co: 0>my fact</co: 0> for a fact from document 0.' }}{{ '<|END_OF_TURN_TOKEN|>' }}{% if add_generation_prompt %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}{% endif %}"
    }
  ]

---

The only reason I ask is I've been running some similar tests and I can't understand how Qwen1.5-110B-Chat scores so highly yet both command-r models score much lower?

I've tried all the qwen-2 and qwen-1.5 models and they all seem to be really bad at "in the style of" which a lot of your benchmarks look to ask for?

I'd love find out exactly what was getting sent as the template for qwen and command-r to see if I can get to the bottom of what is happening!

Just checked the chat-ml template and it is a little odd too:

https://github.com/EQ-bench/EQ-Bench/blob/main_v2_4/instruction-templates/ChatML.yaml

user: user
bot: assistant
turn_template: <|im_start|><|user|>\n<|user-message|><|im_end|>\n<|im_start|><|bot|>\n<|bot-message|><|im_end|>\n
context: |
  <|im_start|>system
  <|system-message|><|im_end|>

Not sure why there is an extra pipe, new line and spaces like that?

[sam-paech]

Great question! And thanks for all the info to help debug.

Ok, so I had a lot of trouble with the command-r models when I was benchmarking them. It's possible that the reason was the extra BOS token, which I guess was likely added since I'm not instantiating the tokenizer with add_special_tokens=False per the discussion you linked.

The way I ran these models for the leaderboard scores is:

c4ai-command-r-v01

EQ-Bench: transformers using the above Cohere template MAGI: lm-eval + transformers

c4ai-command-r-plus

EQ-Bench: either transformers using the above Cohere template, or the huggingface pro api (can't remember which). I just re-ran it with huggingface api and got near-identical result to the leaderboard score. I'm not sure which template they use. MAGI: lm-eval + transformers Creative Writing: huggingface pro api

Qwen-1.5-110B-Chat

EQ-Bench: together.ai api MAGI: lm-eval + transformers Creative Writing: together.ai api

So based on that, and given we don't know what template the huggingface pro api or together.ai is using for these models, I can confidently say that I don't know why it's underperforming! It might be that they are adding an extra BOS token. I don't know how much difference that would make; it might be worth running an A:B comparison with eq-bench / creative writing with the extra BOS token included vs. not.

Otherwise, it's good to be mindful that the creative writing benchmark is pretty subjective. For instance, the Qwen 110B model has a lot of gpt slop which I hate but which sonnet doesn't seem to care about.