#Chat Template for Wizard-Vicuna-13B-Uncensored

14 messages · Page 1 of 1 (latest)

autumn lily
#

I've tried distilling the chat template using the tokenizer_config.json file, but it came up null. What chat template should I use in the Settings for the model Wizard-Vicuna-13B-Uncensored?

clever edgeBOT
#

Hi @autumn lily. I'm kapa.ai, your friendly support bot. I will try to answer your question by searching through the knowledge base. :wave:

#

Based on the information provided in the knowledge sources, I can offer some guidance on finding a chat template for the Wizard-Vicuna-13B-Uncensored model. However, the sources don't provide a specific template for this exact model. Here's what we can deduce:

  1. The model Wizard-Vicuna-13B-Uncensored is mentioned in one of the sources, but it doesn't provide a chat template for it directly.

  2. For models that don't have a built-in chat template or are not in the official model list, you may need to find or create a template yourself.

  3. According to the GPT4All Documentation, if you see a "Clear" option above the chat template field in the Model Settings page instead of a "Reset" option, it means the model doesn't come with a built-in template.

  4. The documentation suggests that in such cases, you should follow the steps to find or create a basic chat template for your model.

To find a suitable chat template, you could try the following steps:

  1. Look for the original model repository on HuggingFace, as suggested in the GPT4All Documentation.

  2. If you can't find a specific template, you might need to create one based on similar models or the general structure of Vicuna models.

  3. As a starting point, you could try a basic template structure like this:

### Human: %1

### Assistant:

This is just a basic example and may need adjustment based on the specific requirements of the Wizard-Vicuna-13B-Uncensored model.

#

If you're still unsure, the GPT4All Documentation suggests looking at existing templates as references and modifying them to suit your model's format. You might also want to check the model's README page or look for community discussions about this specific model for more tailored advice.

Remember, finding the right template might require some trial and error. If you're still having difficulties, you might want to reach out to the GPT4All community for more specific guidance on this particular model.

Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:

#

@autumn lily Kapa.ai is still learning and improving, please let me know how I did by reacting below.

autumn lily
#

@clever edge The app says that this chat template is not in Jinja format.

clever edgeBOT
#

Let me check the knowledge base for you again. One moment! :books:

#

I understand you're encountering an issue with the chat template for the Wizard-Vicuna-13B-Uncensored model. The error message you're seeing about the chat template not being in Jinja format is actually a common issue that was introduced in GPT4All version 3.5. Let me explain this based on the information from our knowledge sources.

According to the GPT4All Documentation, this error occurs when attempting to use an old-style template in GPT4All 3.5 or later versions. The documentation provides steps to resolve this issue:

  1. First, back up your current chat template by copying it to a text file.
  2. In the Model Settings page, look for a "Reset" button for the affected model. If you see it and haven't intentionally modified the prompt template, you can click "Reset".
  3. If you don't see a "Reset" button, click "Clear" instead.
  4. After clearing, you'll need to create a new chat template in Jinja format.

The reason for this change is explained in the GPT4All Documentation. GPT4All v3.5 overhauled the chat template system with three crucial differences:

  1. The chat template now formats an entire conversation instead of a single pair of messages.
  2. It now uses Jinja syntax instead of %1 and %2 placeholders.
  3. The system message should no longer contain control tokens or trailing whitespace.

To create a new Jinja-formatted template, you might start with a basic structure like this:

{% for message in messages %}
{% if message['role'] == 'user' %}
### Human: {{ message['content'] }}
{% elif message['role'] == 'assistant' %}
### Assistant: {{ message['content'] }}
{% endif %}
{% endfor %}
### Assistant:
#

This is a basic example and may need adjustments based on the specific requirements of the Wizard-Vicuna-13B-Uncensored model.

If you're still having difficulties creating a suitable Jinja template, you might want to consult the HuggingFace guide on chat templating, which is recommended in the GPT4All documentation for more advanced template writing tips.

Remember, finding the right template might require some experimentation. If you continue to face issues, consider reaching out to the GPT4All community for more specific guidance on creating a Jinja-formatted template for this particular model.

Note: you can ask me a follow up question by @ mentioning me again :speech_balloon:

#

@autumn lily Kapa.ai is still learning and improving, please let me know how I did by reacting below.

cunning warren
lone shale
#

best template so fare:

cunning warren
#

That Wizar-Vicuna model is probably not finetuned on ChatML syntax, so it might lead to worse results. If you think this is the best jinja template, then you gotta prove it with some benchmarks...