Text Generation
Transformers
PyTorch
llama
text-generation-inference
Inference Endpoints

The prompt style of this model

#18
by Antinis - opened

I would like to confirm the prompt style of the model. The prompt style of the original llama released by meta is like "<s>[INST] user request [\INST] model response", but I noticed that the prompt style of this model is "Q: user request A: model response". So I would like to ask that can't this model use dataset having a prompt style of "[INST]"? The performance seems to decrease after fine-tuning with dataset of this style. Thank you!

this is what our prompt looks like and it works

{%- set ns = namespace(found=false) -%}
{%- for message in messages -%}
    {%- if message['role'] == 'system' -%}
        {%- set ns.found = true -%}
    {%- endif -%}
{%- endfor -%}
{%- if not ns.found -%}
    {{- '' + 'Below is an instruction that describes a task. Write a response that appropriately completes the request.' + '\n\n' -}}
{%- endif %}
{%- for message in messages %}
    {%- if message['role'] == 'system' -%}
        {{- '' + message['content'] + '\n\n' -}}
    {%- else -%}
        {%- if message['role'] == 'user' -%}
            {{-'### Instruction:\n' + message['content'] + '\n\n'-}}
        {%- else -%}
            {{-'### Response:\n' + message['content'] + '\n\n' -}}
        {%- endif -%}
    {%- endif -%}
{%- endfor -%}
{%- if add_generation_prompt -%}
    {{-'### Response:\n'-}}
{%- endif -%}

Sign up or log in to comment