Hf chat template

[tryumanshow]

As far as I know, the apply_chat_template method in HuggingFace isn't currently supported in lm-evaluation-harness. I've addressed this by adding two arguments, is_chat_model and apply_template, to the HFLM class, enabling the use of a predefined template. However, I'd like to note that batchwise operations for apply_chat_template are not yet supported in HuggingFace. Consequently, I've had to resort to iterating through a loop to generate inputs for the model, which consist of dictionaries with keys 'input_ids' and 'attention_mask'. I've heard that you guys are working on adding support for chat models, but I'd submit a pull request just in case. I welcome any feedback you may have!

[LSinev]

I am not sure, but probably these changes also require version bumping of transformers in pyproject.toml (to transformers>=4.35.1 probably, but need check), because chat templates added around 4.34, and were modified later.

[tryumanshow]

Reference

"I'll upgrade the transformers version as you suggested and try again to see if it still works. Thanks for your advice :)

[LSinev]

Mentioning issues with some sort of chat templating requests, hoping subscribed ones to take a peek (and maybe test) at this pull request
#1098
#1209
#1490

[LSinev]

By the way, you can take a peek at previous attempt of some sort of chat templating PR: #1287
Just in case of any pitfalls discussed or mentioned

[eitanturok]

Just wanted to add that it would be really appreciated/helpful to have chatml templates here! This seems like really great work!

[LSinev]

#1560 (comment) this detailed comment may be interesting to readers here too

[PhilipMay]

IMO this is a very important PR. When a model is tested the correct chat template should be applied.
Otherwise it is not a fair comparison.

[djstrong]

Is it a duplicate of #1287?

[eitanturok]

I'd like to highlight that it would be amazing if the lm-evaluation-harness can support multiple templates.

For example, with the new Command-R+ model, you can look at their tokenizer_config file and seee

"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": "{{ 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 }}{{'\n\n## Available Tools\nHere is a list of tools that you have available to you:\n\n'}}{% for tool in tools %}{% if loop.index0 != 0 %}{{ '\n\n'}}{% endif %}{{'```python\ndef ' + tool.name + '('}}{% for param_name, param_fields in tool.parameter_definitions.items() %}{% if loop.index0 != 0 %}{{ ', '}}{% endif %}{{param_name}}: {% if not param_fields.required %}{{'Optional[' + param_fields.type + '] = None'}}{% else %}{{ param_fields.type }}{% endif %}{% endfor %}{{ ') -> List[Dict]:\n    \"\"\"'}}{{ tool.description }}{% if tool.parameter_definitions|length != 0 %}{{ '\n\n    Args:\n        '}}{% for param_name, param_fields in tool.parameter_definitions.items() %}{% if loop.index0 != 0 %}{{ '\n        ' }}{% endif %}{{ param_name + ' ('}}{% if not param_fields.required %}{{'Optional[' + param_fields.type + ']'}}{% else %}{{ param_fields.type }}{% endif %}{{ '): ' + param_fields.description }}{% endfor %}{% endif %}{{ '\n    \"\"\"\n    pass\n```' }}{% endfor %}{{ '<|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|>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|>'}}{% if add_generation_prompt %}{{ '<|START_OF_TURN_TOKEN|><|CHATBOT_TOKEN|>' }}{% endif %}"
    },
    {
      "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 %}"
    }
  ],

This model support different templates for chat, rag, and tool use. If we want to evaluate this model on RAG benchmarks, it is important to apply the RAG template. If we want to evaluate this model on tool use evals, it is important to apply the tool template.

I think in the future there are going to be more custom templates for different use cases. And to eval each model faithfully on that use case it is important to apply the template for that use case.

So I think it would be important to make the lm-evaluation-harness support these different templates for different evals. Would love to hear the thoughts of others on this!

[LSinev]

there are going to be more custom templates

it is important to apply the template for that use case

So, the template name shouldn't be fixed in particular task setup, but, probably, set in model_args part? And also properly logged in results file, I suppose.

[imoneoi]

Any updates to this PR? It's important for evaluating chat/instruction-finetuned models.

[notrichardren]

Based on my understanding, we should prefer #1287 over this change because it only addresses tok_batch_encode which only affects generate_until. Meanwhile, #1287 does chat templating for both generate_until and loglikelihood.