caretech-owl/leo-hessionai-7B-quotes-gguf

Model Card for Model ID

This model is trained to generate german quotes for a given author. The full model can be tested at spaces/caretech-owl/quote-generator-de, here we provide a full model with a 8 bit quantization.

Model Details

Model Description

This fine-tuned model has been trained on the caretech-owl/wikiquote-de-quotes dataset. The model was trained on a prompt like this

prompt_format = "<|im_start|>system\
Dies ist eine Unterhaltung zwischen einem\
intelligenten, hilfsbereitem KI-Assistenten und einem Nutzer.
Der Assistent gibt Antworten in Form von Zitaten.<|im_end|>\n\
<|im_start|>user\
Zitiere {author}<|im_end|>\n<\
|im_start|>assistant\n{quote}<|im_end|>\n"

Where author is itended to be provided by the user, the quote is of format quote + " - " + author. While the model is not able to provide "real" quotes, using authors that are part of the training set and a low temperature for generation results in somewhat realistic quotes that at least sound familiar.

• Developed by: CareTech OWL
• Model type: Causal decoder-only transformer language model
• Language(s) (NLP): German
• License: llama2
• Finetuned from model: LeoLM/leo-hessianai-7b

Uses

import torch
from ctransformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained(
    "caretech-owl/leo-hessionai-7B-quotes-gguf", model_type="llama")

system_prompt = """Dies ist eine Unterhaltung zwischen \
einem intelligenten, hilfsbereitem \
KI-Assistenten und einem Nutzer.
Der Assistent gibt Antworten in Form von Zitaten."""
prompt_format = "<|im_start|>system\n{system_prompt}\
<|im_end|>\n<|im_start|>user\nZitiere {prompt}\
<|im_end|>\n<|im_start|>assistant\n"

def get_quote(author:str, max_new_tokens:int=200):
    query = prompt_format.format(system_prompt=system_prompt, prompt= author)
    output = base_model(query, stop='<|im_end|>', max_new_tokens=max_new_tokens, temperature=0.2, top_k=10)
    print(output)

get_quote("Heinrich Heine")

Training procedure

The following bitsandbytes quantization config was used during training:

• quant_method: gptq
• bits: 8
• tokenizer: None
• dataset: None
• group_size: 32
• damp_percent: 0.1
• desc_act: True
• sym: True
• true_sequential: True
• use_cuda_fp16: False
• model_seqlen: None
• block_name_to_quantize: None
• module_name_preceding_first_block: None
• batch_size: 1
• pad_token_id: None
• use_exllama: True
• max_input_length: None
• exllama_config: {'version': <ExllamaVersion.ONE: 1>}
• cache_block_outputs: True

Framework versions

• PEFT 0.6.2