Politeness Generative Model

Overview

This GPT-based model is a text2text generator that writes a polite version of an input sentence. It is based on gpt-j-6B and was aligned using 29,000 pairs of sentences.

Prompt

You have an input text. Write a polite version of the text preserving the meaning of the input.

Input: What are your thoughts on the proposed merger and its potential effects on our industry?

Output: I'm sorry, but I don't have any thoughts on the proposed merger and its potential effects on our industry.

Quick tutorial

import torch
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer

peft_model_id = "mmendoza/gpt-j-6B-lora-polite-enh"
config = PeftConfig.from_pretrained(peft_model_id)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, return_dict=True, load_in_8bit=True, device_map='auto')
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

Load the Politeness Model

model = PeftModel.from_pretrained(model, peft_model_id)

Prompting

batch = tokenizer("You have an input text. Write a polite version of the text preserving the meaning of the input. 
Input: No card counting allowed in blackjack at the casino. Output: ", return_tensors='pt')

with torch.cuda.amp.autocast():
  output_tokens = model.generate(**batch, max_new_tokens=50, pad_token_id=tokenizer.eos_token_id)

line = tokenizer.decode(output_tokens[0], skip_special_tokens=True)

start = 'Output: '
end = '.'

line = line.replace("\n"," ")
line = (line.split(start))[1].split(end)[0]

"Please refrain from counting cards in blackjack at the casino."

library_name: peft

Training procedure

The following bitsandbytes quantization config was used during training:

load_in_8bit: True
load_in_4bit: False
llm_int8_threshold: 6.0
llm_int8_skip_modules: None
llm_int8_enable_fp32_cpu_offload: False
llm_int8_has_fp16_weight: False
bnb_4bit_quant_type: fp4
bnb_4bit_use_double_quant: False
bnb_4bit_compute_dtype: float32

Framework versions

PEFT 0.4.0.dev0