---
license: mit

datasets:
- samsum

language:
- en

tags:
- summarization
- text-generation
- toxicity-reduction
- reinforcement-learning

widget:
- text: |+
    Summarize the following Conversation: 
    
    Kate: Good morning. 
    Kai: Hi! How official! 
    Kate: I wrote it at 4am 
    Kai: I've noticed. Why? 
    Kate: I had to get up early to catch the bus to the airport 
    Kai: Where are you flying? 
    Kate: To Antwerp! I'm fed up with Cambridge 
    Kai: poor thing. Why? 
    Kate: Just a stupid, elitist place without a soul. Or with a soul made of money. 
    Kai: Try to rest a bit in Belgium, do not work too much. 
    Kate: I have to work, but atleast not in this soulless place. 
    Kai: When are you coming back? 
    Kate: I have to see my supervisor on Monday 
    Kai: not too long a break 
    Kate: Still better than nothing. 
    
    Summary:
  example_title: Summarization Example 1
  
- text: |+
    Summarize the following Conversation: 
    
    Dean: I feel sick 
    Scott: hungover?
    Dean: no, like I ate something bad 
    Scott: what did you eat yesterday? 
    Dean: breakfast at Coffee Lovers
    Scott: this is a rather safe place 
    Dean: and Chinese from TaoTao for dinner 
    Scott: now we have a suspect 
    
    Summary:
  example_title: Summarization Example 2

pipeline_tag: text2text-generation

inference:
  parameters:
    max_new_tokens: 200
    repetition_penalty: 10.5
    top_p: 0.95
    top_k: 50
    temperature: 0.9
    no_repeat_ngram_size: 10
    num_return_sequences: 1
    do_sample: true
---

# Flan-T5 (base-sized) Dialogue Summarization with reduced toxicity using RLAIF
This model is a **two-fold fine-tuned** [Flan-T5 model](https://huggingface.co/google/flan-t5-base) firstly on the [SAMSUM](https://huggingface.co/datasets/samsum) dataset followed by further fine-tuning using **Reinforcement Learning from AI Feedback(RLAIF)** to detoxify model outputs. <br>
Anthropic's Costitutional AI [paper](https://arxiv.org/abs/2212.08073) from 2022, provides some amazing insights on how RLAIF can be leveraged. Do check out if interested!<br>

More specifically, I've fine-tuned this model on a single downstream task of Dialogue Summarization on the above mentioned dataset with a primary objective of reduced toxicity in generated summaries.

## Model description
This Model has the same architecture and Parameters as its base model. Please refer to this [link](https://arxiv.org/abs/2210.11416) to know more about the model details.

## Intended Use & Limitations
This model is intended to summarize the given dialogue in a way that outputs the less toxic summary even when we pass a dialogue that contains toxic phrases or words.<br>
I've fine-tuned the model with an instruction of `Summarize the following Conversation:` that's prepended at the start of each dialogue followed by `Summary: ` keyword at the end that indicates the start of summary.

Note: 
1. The model is primarily trained with an objective of reduced toxicity in the outputs, we can sometimes expect relatively short outputs that might sometimes(rarely) miss the important message in the dialogue but still being true to its primary goal.
2. Currently, HuggingFace doesn't support PEFT model files for Text2Text-Generation Pipeline directly as Hosted Inference API, so please follow the steps mentioned below in the `Usage` section to load and use the model.

## Usage

You can use this model directly to get the summaries:

```python
import torch

from peft import PeftModel, PeftConfig

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer


# Load peft config for pre-trained checkpoint etc.
peft_model_id = "DeathReaper0965/flan-t5-samsum-lora-RLAIF-detoxified"
config = PeftConfig.from_pretrained(peft_model_id)

# load base LLM model and tokenizer
model = AutoModelForSeq2SeqLM.from_pretrained(config.base_model_name_or_path, device_map='auto') # If required, you can add `load_in_8bit=True` for loading model in 8-bit
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)

# Load the Lora model
model = PeftModel.from_pretrained(model, peft_model_id, device_map='auto')

input_ids = tokenizer.encode(
              "Summarize the following Conversation: Dean: I feel sick Scott: hungover? Dean: no, like I ate something bad Scott: what did you eat yesterday? Dean: breakfast at Coffee Lovers' Scott: this is a rather safe place Dean: and Chinese from TaoTao for dinner Scott: now we have a suspect Summary:",
              return_tensors="pt"
            ).to("cuda" if torch.cuda.is_available() else "cpu")

summary = model.generate(
            input_ids = input_ids,
            max_new_tokens=256,
            repetition_penalty=2.5,
            top_p=0.95,
            top_k=50, 
            temperature=0.6,
            no_repeat_ngram_size=2,
            num_return_sequences=1,
            do_sample=True)

output = tokenizer.batch_decode(summary, skip_special_tokens=True)

###########OUTPUT###########
# "Dean ate breakfast at Coffee Lovers' yesterday and Chinese from TaoTao for dinner."
```

> Designed and Developed with <span style="color: #e25555;">&hearts;</span> by [Praneet](https://deathreaper0965.github.io/) | [LinkedIn](http://linkedin.com/in/deathreaper0965) | [GitHub](https://github.com/DeathReaper0965/)