tweeteval_llama2-70b-chat_hf-inference-api.yaml

# Config for everything related to dataset generation inputs
input_config:
    # Dataset metadata
    domain: tweets
    language: en

    # Dataset generator parameters
    quantity: 10

    # HuggingFace dataset params
    dataset: tweet_eval
    dataset_text_column: text
    dataset_params:
        split: train
        name: irony
    
    # Prompt template
    template: >- 
        Write a tweet response to the following tweet:\n'{text}'\nResponse: 
    
    # Extractor params
    extractor: auxiliary
    max_input_tokens: 256

# Config for model instantiation
model_config:
    provider: hf_remote
    model_name: meta-llama/Llama-2-70b-chat-hf
    url: https://api-inference.huggingface.co/models/meta-llama/Llama-2-70b-chat-hf
    max_retries: 5
    threads: 8

# Decoding args
generation_config:
    # Ignore `max_new_tokens` or `min_new_tokens` to get automatic length estimation
    # min_new_tokens: 5
    # max_new_tokens: 10
    do_sample: true
    top_p: 0.9
    temperature: 0.6
    repetition_penalty: 1.0
    return_full_text: false