Coursera
Fine-Tune FLAN-T5 with Reinforcement Learning (PPO) and PEFT to Generate Less-Toxic Summaries
In this notebook, you will fine-tune a FLAN-T5 model to generate less toxic content with Meta AI’s hate speech reward model. The reward model is a binary classifier that predicts either “not hate” or “hate” for the given text. You will use Proximal Policy Optimization (PPO) to fine-tune and reduce the model’s toxicity.
Table of Contents
- 1 - Set up Kernel and Required Dependencies
- 2 - Load FLAN-T5 Model, Prepare Reward Model and Toxicity Evaluator
- 3 - Perform Fine-Tuning to Detoxify the Summaries
1 - Set up Kernel and Required Dependencies
First, check that the correct kernel is chosen.
You can click on that (top right of the screen) to see and check the details of the image, kernel, and instance type.
Now install the required packages to use PyTorch and Hugging Face transformers and datasets.
%pip install --upgrade pip
%pip install --disable-pip-version-check \
torch==1.13.1 \
torchdata==0.5.1 --quiet
%pip install \
transformers==4.27.2 \
datasets==2.11.0 \
evaluate==0.4.0 \
rouge_score==0.1.2 \
peft==0.3.0 --quiet
# Installing the Reinforcement Learning library directly from github.
%pip install git+https://github.com/lvwerra/trl.git@25fa1bd
Requirement already satisfied: pip in /opt/conda/lib/python3.7/site-packages (23.2.1)
Collecting pip
Obtaining dependency information for pip from https://files.pythonhosted.org/packages/e0/63/b428aaca15fcd98c39b07ca7149e24bc14205ad0f1c80ba2b01835aedde1/pip-23.3-py3-none-any.whl.metadata
Downloading pip-23.3-py3-none-any.whl.metadata (3.5 kB)
Downloading pip-23.3-py3-none-any.whl (2.1 MB)
[2K [90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━[0m [32m2.1/2.1 MB[0m [31m19.8 MB/s[0m eta [36m0:00:00[0m:00:01[0m
[?25h[33mDEPRECATION: pyodbc 4.0.0-unsupported has a non-standard version number. pip 23.3 will enforce this behaviour change. A possible replacement is to upgrade to a newer version of pyodbc or contact the author to suggest that they release a version with a conforming version number. Discussion can be found at https://github.com/pypa/pip/issues/12063[0m[33m
[0mInstalling collected packages: pip
Attempting uninstall: pip
Found existing installation: pip 23.2.1
Uninstalling pip-23.2.1:
Successfully uninstalled pip-23.2.1
Successfully installed pip-23.3
[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv[0m[33m
[0mNote: you may need to restart the kernel to use updated packages.
[33mDEPRECATION: pyodbc 4.0.0-unsupported has a non-standard version number. pip 24.0 will enforce this behaviour change. A possible replacement is to upgrade to a newer version of pyodbc or contact the author to suggest that they release a version with a conforming version number. Discussion can be found at https://github.com/pypa/pip/issues/12063[0m[33m
[0m[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv[0m[33m
[0mNote: you may need to restart the kernel to use updated packages.
[33mDEPRECATION: pyodbc 4.0.0-unsupported has a non-standard version number. pip 24.0 will enforce this behaviour change. A possible replacement is to upgrade to a newer version of pyodbc or contact the author to suggest that they release a version with a conforming version number. Discussion can be found at https://github.com/pypa/pip/issues/12063[0m[33m
[0m[31mERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
pytest-astropy 0.8.0 requires pytest-cov>=2.0, which is not installed.
pytest-astropy 0.8.0 requires pytest-filter-subpackage>=0.1, which is not installed.
spyder 4.0.1 requires pyqt5<5.13; python_version >= "3", which is not installed.
spyder 4.0.1 requires pyqtwebengine<5.13; python_version >= "3", which is not installed.
notebook 6.5.5 requires pyzmq<25,>=17, but you have pyzmq 25.1.1 which is incompatible.
pathos 0.3.1 requires dill>=0.3.7, but you have dill 0.3.6 which is incompatible.
pathos 0.3.1 requires multiprocess>=0.70.15, but you have multiprocess 0.70.14 which is incompatible.
sparkmagic 0.20.4 requires nest-asyncio==1.5.5, but you have nest-asyncio 1.5.7 which is incompatible.
spyder 4.0.1 requires jedi==0.14.1, but you have jedi 0.19.0 which is incompatible.[0m[31m
[0m[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv[0m[33m
[0m[33mWARNING: There was an error checking the latest version of pip.[0m[33m
[0mNote: you may need to restart the kernel to use updated packages.
Collecting git+https://github.com/lvwerra/trl.git@25fa1bd
Cloning https://github.com/lvwerra/trl.git (to revision 25fa1bd) to /tmp/pip-req-build-n33yd2ny
Running command git clone --filter=blob:none --quiet https://github.com/lvwerra/trl.git /tmp/pip-req-build-n33yd2ny
[33m WARNING: Did not find branch or tag '25fa1bd', assuming revision or ref.[0m[33m
[0m Running command git checkout -q 25fa1bd
Resolved https://github.com/lvwerra/trl.git to commit 25fa1bd
Preparing metadata (setup.py) ... [?25ldone
[?25hRequirement already satisfied: torch>=1.4.0 in /opt/conda/lib/python3.7/site-packages (from trl==0.4.2.dev0) (1.13.1)
Requirement already satisfied: transformers>=4.18.0 in /opt/conda/lib/python3.7/site-packages (from trl==0.4.2.dev0) (4.27.2)
Requirement already satisfied: numpy>=1.18.2 in /opt/conda/lib/python3.7/site-packages (from trl==0.4.2.dev0) (1.21.6)
Requirement already satisfied: accelerate in /opt/conda/lib/python3.7/site-packages (from trl==0.4.2.dev0) (0.20.3)
Requirement already satisfied: datasets in /opt/conda/lib/python3.7/site-packages (from trl==0.4.2.dev0) (2.11.0)
Requirement already satisfied: typing-extensions in /opt/conda/lib/python3.7/site-packages (from torch>=1.4.0->trl==0.4.2.dev0) (4.7.1)
Requirement already satisfied: nvidia-cuda-runtime-cu11==11.7.99 in /opt/conda/lib/python3.7/site-packages (from torch>=1.4.0->trl==0.4.2.dev0) (11.7.99)
Requirement already satisfied: nvidia-cudnn-cu11==8.5.0.96 in /opt/conda/lib/python3.7/site-packages (from torch>=1.4.0->trl==0.4.2.dev0) (8.5.0.96)
Requirement already satisfied: nvidia-cublas-cu11==11.10.3.66 in /opt/conda/lib/python3.7/site-packages (from torch>=1.4.0->trl==0.4.2.dev0) (11.10.3.66)
Requirement already satisfied: nvidia-cuda-nvrtc-cu11==11.7.99 in /opt/conda/lib/python3.7/site-packages (from torch>=1.4.0->trl==0.4.2.dev0) (11.7.99)
Requirement already satisfied: setuptools in /opt/conda/lib/python3.7/site-packages (from nvidia-cublas-cu11==11.10.3.66->torch>=1.4.0->trl==0.4.2.dev0) (65.5.1)
Requirement already satisfied: wheel in /opt/conda/lib/python3.7/site-packages (from nvidia-cublas-cu11==11.10.3.66->torch>=1.4.0->trl==0.4.2.dev0) (0.41.1)
Requirement already satisfied: filelock in /opt/conda/lib/python3.7/site-packages (from transformers>=4.18.0->trl==0.4.2.dev0) (3.0.12)
Requirement already satisfied: huggingface-hub<1.0,>=0.11.0 in /opt/conda/lib/python3.7/site-packages (from transformers>=4.18.0->trl==0.4.2.dev0) (0.16.4)
Requirement already satisfied: packaging>=20.0 in /opt/conda/lib/python3.7/site-packages (from transformers>=4.18.0->trl==0.4.2.dev0) (23.2)
Requirement already satisfied: pyyaml>=5.1 in /opt/conda/lib/python3.7/site-packages (from transformers>=4.18.0->trl==0.4.2.dev0) (6.0.1)
Requirement already satisfied: regex!=2019.12.17 in /opt/conda/lib/python3.7/site-packages (from transformers>=4.18.0->trl==0.4.2.dev0) (2023.8.8)
Requirement already satisfied: requests in /opt/conda/lib/python3.7/site-packages (from transformers>=4.18.0->trl==0.4.2.dev0) (2.31.0)
Requirement already satisfied: tokenizers!=0.11.3,<0.14,>=0.11.1 in /opt/conda/lib/python3.7/site-packages (from transformers>=4.18.0->trl==0.4.2.dev0) (0.13.3)
Requirement already satisfied: tqdm>=4.27 in /opt/conda/lib/python3.7/site-packages (from transformers>=4.18.0->trl==0.4.2.dev0) (4.66.1)
Requirement already satisfied: importlib-metadata in /opt/conda/lib/python3.7/site-packages (from transformers>=4.18.0->trl==0.4.2.dev0) (6.7.0)
Requirement already satisfied: psutil in /opt/conda/lib/python3.7/site-packages (from accelerate->trl==0.4.2.dev0) (5.6.7)
Requirement already satisfied: pyarrow>=8.0.0 in /opt/conda/lib/python3.7/site-packages (from datasets->trl==0.4.2.dev0) (12.0.1)
Requirement already satisfied: dill<0.3.7,>=0.3.0 in /opt/conda/lib/python3.7/site-packages (from datasets->trl==0.4.2.dev0) (0.3.6)
Requirement already satisfied: pandas in /opt/conda/lib/python3.7/site-packages (from datasets->trl==0.4.2.dev0) (1.3.5)
Requirement already satisfied: xxhash in /opt/conda/lib/python3.7/site-packages (from datasets->trl==0.4.2.dev0) (3.4.1)
Requirement already satisfied: multiprocess in /opt/conda/lib/python3.7/site-packages (from datasets->trl==0.4.2.dev0) (0.70.14)
Requirement already satisfied: fsspec>=2021.11.1 in /opt/conda/lib/python3.7/site-packages (from fsspec[http]>=2021.11.1->datasets->trl==0.4.2.dev0) (2023.1.0)
Requirement already satisfied: aiohttp in /opt/conda/lib/python3.7/site-packages (from datasets->trl==0.4.2.dev0) (3.8.5)
Requirement already satisfied: responses<0.19 in /opt/conda/lib/python3.7/site-packages (from datasets->trl==0.4.2.dev0) (0.18.0)
Requirement already satisfied: attrs>=17.3.0 in /opt/conda/lib/python3.7/site-packages (from aiohttp->datasets->trl==0.4.2.dev0) (23.1.0)
Requirement already satisfied: charset-normalizer<4.0,>=2.0 in /opt/conda/lib/python3.7/site-packages (from aiohttp->datasets->trl==0.4.2.dev0) (2.0.4)
Requirement already satisfied: multidict<7.0,>=4.5 in /opt/conda/lib/python3.7/site-packages (from aiohttp->datasets->trl==0.4.2.dev0) (6.0.4)
Requirement already satisfied: async-timeout<5.0,>=4.0.0a3 in /opt/conda/lib/python3.7/site-packages (from aiohttp->datasets->trl==0.4.2.dev0) (4.0.3)
Requirement already satisfied: yarl<2.0,>=1.0 in /opt/conda/lib/python3.7/site-packages (from aiohttp->datasets->trl==0.4.2.dev0) (1.9.2)
Requirement already satisfied: frozenlist>=1.1.1 in /opt/conda/lib/python3.7/site-packages (from aiohttp->datasets->trl==0.4.2.dev0) (1.3.3)
Requirement already satisfied: aiosignal>=1.1.2 in /opt/conda/lib/python3.7/site-packages (from aiohttp->datasets->trl==0.4.2.dev0) (1.3.1)
Requirement already satisfied: asynctest==0.13.0 in /opt/conda/lib/python3.7/site-packages (from aiohttp->datasets->trl==0.4.2.dev0) (0.13.0)
Requirement already satisfied: idna<4,>=2.5 in /opt/conda/lib/python3.7/site-packages (from requests->transformers>=4.18.0->trl==0.4.2.dev0) (2.8)
Requirement already satisfied: urllib3<3,>=1.21.1 in /opt/conda/lib/python3.7/site-packages (from requests->transformers>=4.18.0->trl==0.4.2.dev0) (2.0.4)
Requirement already satisfied: certifi>=2017.4.17 in /opt/conda/lib/python3.7/site-packages (from requests->transformers>=4.18.0->trl==0.4.2.dev0) (2023.7.22)
Requirement already satisfied: zipp>=0.5 in /opt/conda/lib/python3.7/site-packages (from importlib-metadata->transformers>=4.18.0->trl==0.4.2.dev0) (2.2.0)
Requirement already satisfied: python-dateutil>=2.7.3 in /opt/conda/lib/python3.7/site-packages (from pandas->datasets->trl==0.4.2.dev0) (2.8.2)
Requirement already satisfied: pytz>=2017.3 in /opt/conda/lib/python3.7/site-packages (from pandas->datasets->trl==0.4.2.dev0) (2019.3)
Requirement already satisfied: six>=1.5 in /opt/conda/lib/python3.7/site-packages (from python-dateutil>=2.7.3->pandas->datasets->trl==0.4.2.dev0) (1.14.0)
Building wheels for collected packages: trl
Building wheel for trl (setup.py) ... [?25ldone
[?25h Created wheel for trl: filename=trl-0.4.2.dev0-py3-none-any.whl size=67532 sha256=b1db3482276cde915e86f5eb0e6f699196fb760ff8ded5d2111d098b01ccf42a
Stored in directory: /tmp/pip-ephem-wheel-cache-cmq48obh/wheels/41/26/75/08a45cee1a1bba06c4f340451483cdfe150f4c8dad3876fb2e
Successfully built trl
[33mDEPRECATION: pyodbc 4.0.0-unsupported has a non-standard version number. pip 24.0 will enforce this behaviour change. A possible replacement is to upgrade to a newer version of pyodbc or contact the author to suggest that they release a version with a conforming version number. Discussion can be found at https://github.com/pypa/pip/issues/12063[0m[33m
[0mInstalling collected packages: trl
Successfully installed trl-0.4.2.dev0
[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv[0m[33m
[0m[33mWARNING: There was an error checking the latest version of pip.[0m[33m
[0mNote: you may need to restart the kernel to use updated packages.
Import the necessary components. Some of them are new for this week, they will be discussed later in the notebook.
from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification, AutoModelForSeq2SeqLM, GenerationConfig
from datasets import load_dataset
from peft import PeftModel, PeftConfig, LoraConfig, TaskType
# trl: Transformer Reinforcement Learning library
from trl import PPOTrainer, PPOConfig, AutoModelForSeq2SeqLMWithValueHead
from trl import create_reference_model
from trl.core import LengthSampler
import torch
import evaluate
import numpy as np
import pandas as pd
# tqdm library makes the loops show a smart progress meter.
from tqdm import tqdm
tqdm.pandas()
2 - Load FLAN-T5 Model, Prepare Reward Model and Toxicity Evaluator
2.1 - Load Data and FLAN-T5 Model Fine-Tuned with Summarization Instruction
You will keep working with the same Hugging Face dataset DialogSum and the pre-trained model FLAN-T5.
model_name="google/flan-t5-base"
huggingface_dataset_name = "knkarthick/dialogsum"
dataset_original = load_dataset(huggingface_dataset_name)
dataset_original
Downloading readme: 0%| | 0.00/4.65k [00:00<?, ?B/s]
Downloading and preparing dataset csv/knkarthick--dialogsum to /root/.cache/huggingface/datasets/knkarthick___csv/knkarthick--dialogsum-3005b557c2c04c1d/0.0.0/6954658bab30a358235fa864b05cf819af0e179325c740e4bc853bcc7ec513e1...
Downloading data files: 0%| | 0/3 [00:00<?, ?it/s]
Downloading data: 0%| | 0.00/11.3M [00:00<?, ?B/s]
Downloading data: 0%| | 0.00/1.35M [00:00<?, ?B/s]
Downloading data: 0%| | 0.00/442k [00:00<?, ?B/s]
Extracting data files: 0%| | 0/3 [00:00<?, ?it/s]
Generating train split: 0 examples [00:00, ? examples/s]
Generating test split: 0 examples [00:00, ? examples/s]
Generating validation split: 0 examples [00:00, ? examples/s]
Dataset csv downloaded and prepared to /root/.cache/huggingface/datasets/knkarthick___csv/knkarthick--dialogsum-3005b557c2c04c1d/0.0.0/6954658bab30a358235fa864b05cf819af0e179325c740e4bc853bcc7ec513e1. Subsequent calls will reuse this data.
0%| | 0/3 [00:00<?, ?it/s]
DatasetDict({
train: Dataset({
features: ['id', 'dialogue', 'summary', 'topic'],
num_rows: 12460
})
test: Dataset({
features: ['id', 'dialogue', 'summary', 'topic'],
num_rows: 1500
})
validation: Dataset({
features: ['id', 'dialogue', 'summary', 'topic'],
num_rows: 500
})
})
The next step will be to preprocess the dataset. You will take only a part of it, then filter the dialogues of a particular length (just to make those examples long enough and, at the same time, easy to read). Then wrap each dialogue with the instruction and tokenize the prompts. Save the token ids in the field input_ids and decoded version of the prompts in the field query.
You could do that all step by step in the cell below, but it is a good habit to organize that all in a function build_dataset:
def build_dataset(model_name,
dataset_name,
input_min_text_length,
input_max_text_length):
"""
Preprocess the dataset and split it into train and test parts.
Parameters:
- model_name (str): Tokenizer model name.
- dataset_name (str): Name of the dataset to load.
- input_min_text_length (int): Minimum length of the dialogues.
- input_max_text_length (int): Maximum length of the dialogues.
Returns:
- dataset_splits (datasets.dataset_dict.DatasetDict): Preprocessed dataset containing train and test parts.
"""
# load dataset (only "train" part will be enough for this lab).
dataset = load_dataset(dataset_name, split="train")
# Filter the dialogues of length between input_min_text_length and input_max_text_length characters.
dataset = dataset.filter(lambda x: len(x["dialogue"]) > input_min_text_length and len(x["dialogue"]) <= input_max_text_length, batched=False)
# Prepare tokenizer. Setting device_map="auto" allows to switch between GPU and CPU automatically.
tokenizer = AutoTokenizer.from_pretrained(model_name, device_map="auto")
def tokenize(sample):
# Wrap each dialogue with the instruction.
prompt = f"""
Summarize the following conversation.
{sample["dialogue"]}
Summary:
"""
sample["input_ids"] = tokenizer.encode(prompt)
# This must be called "query", which is a requirement of our PPO library.
sample["query"] = tokenizer.decode(sample["input_ids"])
return sample
# Tokenize each dialogue.
dataset = dataset.map(tokenize, batched=False)
dataset.set_format(type="torch")
# Split the dataset into train and test parts.
dataset_splits = dataset.train_test_split(test_size=0.2, shuffle=False, seed=42)
return dataset_splits
dataset = build_dataset(model_name=model_name,
dataset_name=huggingface_dataset_name,
input_min_text_length=200,
input_max_text_length=1000)
print(dataset)
Found cached dataset csv (/root/.cache/huggingface/datasets/knkarthick___csv/knkarthick--dialogsum-3005b557c2c04c1d/0.0.0/6954658bab30a358235fa864b05cf819af0e179325c740e4bc853bcc7ec513e1)
Filter: 0%| | 0/12460 [00:00<?, ? examples/s]
Downloading (…)okenizer_config.json: 0%| | 0.00/2.54k [00:00<?, ?B/s]
Downloading spiece.model: 0%| | 0.00/792k [00:00<?, ?B/s]
Downloading (…)/main/tokenizer.json: 0%| | 0.00/2.42M [00:00<?, ?B/s]
Downloading (…)cial_tokens_map.json: 0%| | 0.00/2.20k [00:00<?, ?B/s]
Map: 0%| | 0/10022 [00:00<?, ? examples/s]
DatasetDict({
train: Dataset({
features: ['id', 'dialogue', 'summary', 'topic', 'input_ids', 'query'],
num_rows: 8017
})
test: Dataset({
features: ['id', 'dialogue', 'summary', 'topic', 'input_ids', 'query'],
num_rows: 2005
})
})
In the previous lab, you fine-tuned the PEFT model with summarization instructions. The training in the notebook was done on a subset of data. Then you downloaded the checkpoint of the fully trained PEFT model from S3.
Let’s load the same model checkpoint here:
!aws s3 cp --recursive s3://dlai-generative-ai/models/peft-dialogue-summary-checkpoint/ ./peft-dialogue-summary-checkpoint-from-s3/
download: s3://dlai-generative-ai/models/peft-dialogue-summary-checkpoint/special_tokens_map.json to peft-dialogue-summary-checkpoint-from-s3/special_tokens_map.json
download: s3://dlai-generative-ai/models/peft-dialogue-summary-checkpoint/adapter_config.json to peft-dialogue-summary-checkpoint-from-s3/adapter_config.json
download: s3://dlai-generative-ai/models/peft-dialogue-summary-checkpoint/tokenizer_config.json to peft-dialogue-summary-checkpoint-from-s3/tokenizer_config.json
download: s3://dlai-generative-ai/models/peft-dialogue-summary-checkpoint/tokenizer.json to peft-dialogue-summary-checkpoint-from-s3/tokenizer.json
download: s3://dlai-generative-ai/models/peft-dialogue-summary-checkpoint/adapter_model.bin to peft-dialogue-summary-checkpoint-from-s3/adapter_model.bin
List the model item and check its size (it’s less than 15 Mb):
!ls -alh ./peft-dialogue-summary-checkpoint-from-s3/adapter_model.bin
-rw-r--r-- 1 root root 14M May 15 11:18 ./peft-dialogue-summary-checkpoint-from-s3/adapter_model.bin
Prepare a function to pull out the number of model parameters (it is the same as in the previous lab):
def print_number_of_trainable_model_parameters(model):
trainable_model_params = 0
all_model_params = 0
for _, param in model.named_parameters():
all_model_params += param.numel()
if param.requires_grad:
trainable_model_params += param.numel()
return f"\ntrainable model parameters: {trainable_model_params}\nall model parameters: {all_model_params}\npercentage of trainable model parameters: {100 * trainable_model_params / all_model_params:.2f}%"
Add the adapter to the original FLAN-T5 model. In the previous lab you were adding the fully trained adapter only for inferences, so there was no need to pass LoRA configurations doing that. Now you need to pass them to the constructed PEFT model, also putting is_trainable=True.
lora_config = LoraConfig(
r=32, # Rank
lora_alpha=32,
target_modules=["q", "v"],
lora_dropout=0.05,
bias="none",
task_type=TaskType.SEQ_2_SEQ_LM # FLAN-T5
)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name,
torch_dtype=torch.bfloat16)
peft_model = PeftModel.from_pretrained(model,
'./peft-dialogue-summary-checkpoint-from-s3/',
lora_config=lora_config,
torch_dtype=torch.bfloat16,
device_map="auto",
is_trainable=True)
print(f'PEFT model parameters to be updated:\n{print_number_of_trainable_model_parameters(peft_model)}\n')
Downloading (…)lve/main/config.json: 0%| | 0.00/1.40k [00:00<?, ?B/s]
Downloading pytorch_model.bin: 0%| | 0.00/990M [00:00<?, ?B/s]
Downloading (…)neration_config.json: 0%| | 0.00/147 [00:00<?, ?B/s]
PEFT model parameters to be updated:
trainable model parameters: 3538944
all model parameters: 251116800
percentage of trainable model parameters: 1.41%
In this lab, you are preparing to fine-tune the LLM using Reinforcement Learning (RL). RL will be briefly discussed in the next section of this lab, but at this stage, you just need to prepare the Proximal Policy Optimization (PPO) model passing the instruct-fine-tuned PEFT model to it. PPO will be used to optimize the RL policy against the reward model.
ppo_model = AutoModelForSeq2SeqLMWithValueHead.from_pretrained(peft_model,
torch_dtype=torch.bfloat16,
is_trainable=True)
print(f'PPO model parameters to be updated (ValueHead + 769 params):\n{print_number_of_trainable_model_parameters(ppo_model)}\n')
print(ppo_model.v_head)
PPO model parameters to be updated (ValueHead + 769 params):
trainable model parameters: 3539713
all model parameters: 251117569
percentage of trainable model parameters: 1.41%
ValueHead(
(dropout): Dropout(p=0.1, inplace=False)
(summary): Linear(in_features=768, out_features=1, bias=True)
(flatten): Flatten(start_dim=1, end_dim=-1)
)
During PPO, only a few parameters will be updated. Specifically, the parameters of the ValueHead. More information about this class of models can be found in the documentation. The number of trainable parameters can be computed as $(n+1)*m$, where $n$ is the number of input units (here $n=768$) and $m$ is the number of output units (you have $m=1$). The $+1$ term in the equation takes into account the bias term.
Now create a frozen copy of the PPO which will not be fine-tuned - a reference model. The reference model will represent the LLM before detoxification. None of the parameters of the reference model will be updated during PPO training. This is on purpose.
ref_model = create_reference_model(ppo_model)
print(f'Reference model parameters to be updated:\n{print_number_of_trainable_model_parameters(ref_model)}\n')
Reference model parameters to be updated:
trainable model parameters: 0
all model parameters: 251117569
percentage of trainable model parameters: 0.00%
Everything is set. It is time to prepare the reward model!
2.2 - Prepare Reward Model
Reinforcement Learning (RL) is one type of machine learning where agents take actions in an environment aimed at maximizing their cumulative rewards. The agent’s behavior is defined by the policy. And the goal of reinforcement learning is for the agent to learn an optimal, or nearly-optimal, policy that maximizes the reward function.
In the previous section the original policy is based on the instruct PEFT model - this is the LLM before detoxification. Then you could ask human labelers to give feedback on the outputs’ toxicity. However, it can be expensive to use them for the entire fine-tuning process. A practical way to avoid that is to use a reward model encouraging the agent to detoxify the dialogue summaries. The intuitive approach would be to do some form of sentiment analysis across two classes (nothate and hate) and give a higher reward if there is higher a chance of getting class nothate as an output.
For example, we can mention that having human labelers for the entire finetuning process can be expensive. A practical way to avoid that is to use a reward model.
use feedback generated by a model
You will use Meta AI’s RoBERTa-based hate speech model for the reward model. This model will output logits and then predict probabilities across two classes: nothate and hate. The logits of the output nothate will be taken as a positive reward. Then, the model will be fine-tuned with PPO using those reward values.
Create the instance of the required model class for the RoBERTa model. You also need to load a tokenizer to test the model. Notice that the model label 0 will correspond to the class nothate and label 1 to the class hate.
toxicity_model_name = "facebook/roberta-hate-speech-dynabench-r4-target"
toxicity_tokenizer = AutoTokenizer.from_pretrained(toxicity_model_name, device_map="auto")
toxicity_model = AutoModelForSequenceClassification.from_pretrained(toxicity_model_name, device_map="auto")
print(toxicity_model.config.id2label)
Downloading (…)okenizer_config.json: 0%| | 0.00/1.11k [00:00<?, ?B/s]
Downloading (…)olve/main/vocab.json: 0%| | 0.00/899k [00:00<?, ?B/s]
Downloading (…)olve/main/merges.txt: 0%| | 0.00/456k [00:00<?, ?B/s]
Downloading (…)cial_tokens_map.json: 0%| | 0.00/239 [00:00<?, ?B/s]
Downloading (…)lve/main/config.json: 0%| | 0.00/816 [00:00<?, ?B/s]
Downloading pytorch_model.bin: 0%| | 0.00/499M [00:00<?, ?B/s]
The model weights are not tied. Please use the `tie_weights` method before using the `infer_auto_device` function.
{0: 'nothate', 1: 'hate'}
Take some non-toxic text, tokenize it, and pass it to the model. Print the output logits, probabilities, and the corresponding reward that will be used for fine-tuning.
non_toxic_text = "#Person 1# tells Tommy that he didn't like the movie."
toxicity_input_ids = toxicity_tokenizer(non_toxic_text, return_tensors="pt").input_ids
logits = toxicity_model(input_ids=toxicity_input_ids).logits
print(f'logits [not hate, hate]: {logits.tolist()[0]}')
# Print the probabilities for [not hate, hate]
probabilities = logits.softmax(dim=-1).tolist()[0]
print(f'probabilities [not hate, hate]: {probabilities}')
# get the logits for "not hate" - this is the reward!
not_hate_index = 0
nothate_reward = (logits[:, not_hate_index]).tolist()
print(f'reward (high): {nothate_reward}')
logits [not hate, hate]: [3.114100694656372, -2.4896175861358643]
probabilities [not hate, hate]: [0.9963293671607971, 0.003670616541057825]
reward (high): [3.114100694656372]
Let’s show a toxic comment. This will have a low reward because it is more toxic.
toxic_text = "#Person 1# tells Tommy that the movie was terrible, dumb and stupid."
toxicity_input_ids = toxicity_tokenizer(toxic_text, return_tensors="pt").input_ids
logits = toxicity_model(toxicity_input_ids).logits
print(f'logits [not hate, hate]: {logits.tolist()[0]}')
# Print the probabilities for [not hate, hate]
probabilities = logits.softmax(dim=-1).tolist()[0]
print(f'probabilities [not hate, hate]: {probabilities}')
# Get the logits for "not hate" - this is the reward!
nothate_reward = (logits[:, not_hate_index]).tolist()
print(f'reward (low): {nothate_reward}')
logits [not hate, hate]: [-0.6921188831329346, 0.3722729980945587]
probabilities [not hate, hate]: [0.25647106766700745, 0.7435289621353149]
reward (low): [-0.6921188831329346]
Setup Hugging Face inference pipeline to simplify the code for the toxicity reward model:
device = 0 if torch.cuda.is_available() else "cpu"
sentiment_pipe = pipeline("sentiment-analysis",
model=toxicity_model_name,
device=device)
reward_logits_kwargs = {
"top_k": None, # Return all scores.
"function_to_apply": "none", # Set to "none" to retrieve raw logits.
"batch_size": 16
}
reward_probabilities_kwargs = {
"top_k": None, # Return all scores.
"function_to_apply": "softmax", # Set to "softmax" to apply softmax and retrieve probabilities.
"batch_size": 16
}
print("Reward model output:")
print("For non-toxic text")
print(sentiment_pipe(non_toxic_text, **reward_logits_kwargs))
print(sentiment_pipe(non_toxic_text, **reward_probabilities_kwargs))
print("For toxic text")
print(sentiment_pipe(toxic_text, **reward_logits_kwargs))
print(sentiment_pipe(toxic_text, **reward_probabilities_kwargs))
Reward model output:
For non-toxic text
[{'label': 'nothate', 'score': 3.114100694656372}, {'label': 'hate', 'score': -2.4896175861358643}]
[{'label': 'nothate', 'score': 0.9963293671607971}, {'label': 'hate', 'score': 0.003670616541057825}]
For toxic text
[{'label': 'hate', 'score': 0.3722729980945587}, {'label': 'nothate', 'score': -0.6921188831329346}]
[{'label': 'hate', 'score': 0.7435289621353149}, {'label': 'nothate', 'score': 0.25647106766700745}]
The outputs are the logits for both nothate (positive) and hate (negative) classes. But PPO will be using logits only of the nothate class as the positive reward signal used to help detoxify the LLM outputs.
print(sentiment_pipe(non_toxic_text, **reward_logits_kwargs))
print(sentiment_pipe(non_toxic_text, **reward_probabilities_kwargs))
[{'label': 'nothate', 'score': 3.114100694656372}, {'label': 'hate', 'score': -2.4896175861358643}]
[{'label': 'nothate', 'score': 0.9963293671607971}, {'label': 'hate', 'score': 0.003670616541057825}]
print(sentiment_pipe(toxic_text, **reward_logits_kwargs))
print(sentiment_pipe(toxic_text, **reward_probabilities_kwargs))
[{'label': 'hate', 'score': 0.3722729980945587}, {'label': 'nothate', 'score': -0.6921188831329346}]
[{'label': 'hate', 'score': 0.7435289621353149}, {'label': 'nothate', 'score': 0.25647106766700745}]
2.3 - Evaluate Toxicity
To evaluate the model before and after fine-tuning/detoxification you need to set up the toxicity evaluation metric. The toxicity score is a decimal value between 0 and 1 where 1 is the highest toxicity.
toxicity_evaluator = evaluate.load("toxicity",
toxicity_model_name,
module_type="measurement",
toxic_label="hate")
Downloading builder script: 0%| | 0.00/6.08k [00:00<?, ?B/s]
Try to calculate toxicity for the same sentences as in section 2.2. It’s no surprise that the toxicity scores are the probabilities of hate class returned directly from the reward model.
toxicity_score = toxicity_evaluator.compute(predictions=[
non_toxic_text
])
print("Toxicity score for non-toxic text:")
print(toxicity_score["toxicity"])
toxicity_score = toxicity_evaluator.compute(predictions=[
toxic_text
])
print("\nToxicity score for toxic text:")
print(toxicity_score["toxicity"])
Toxicity score for non-toxic text:
[0.003670616541057825]
Toxicity score for toxic text:
[0.7435289621353149]
This evaluator can be used to compute the toxicity of the dialogues prepared in section 2.1. You will need to pass the test dataset (dataset["test"]), the same tokenizer which was used in that section, the frozen PEFT model prepared in section 2.2, and the toxicity evaluator. It is convenient to wrap the required steps in the function evaluate_toxicity.
def evaluate_toxicity(model,
toxicity_evaluator,
tokenizer,
dataset,
num_samples):
"""
Preprocess the dataset and split it into train and test parts.
Parameters:
- model (trl model): Model to be evaluated.
- toxicity_evaluator (evaluate_modules toxicity metrics): Toxicity evaluator.
- tokenizer (transformers tokenizer): Tokenizer to be used.
- dataset (dataset): Input dataset for the evaluation.
- num_samples (int): Maximum number of samples for the evaluation.
Returns:
tuple: A tuple containing two numpy.float64 values:
- mean (numpy.float64): Mean of the samples toxicity.
- std (numpy.float64): Standard deviation of the samples toxicity.
"""
max_new_tokens=100
toxicities = []
input_texts = []
for i, sample in tqdm(enumerate(dataset)):
input_text = sample["query"]
if i > num_samples:
break
input_ids = tokenizer(input_text, return_tensors="pt", padding=True).input_ids
generation_config = GenerationConfig(max_new_tokens=max_new_tokens,
top_k=0.0,
top_p=1.0,
do_sample=True)
response_token_ids = model.generate(input_ids=input_ids,
generation_config=generation_config)
generated_text = tokenizer.decode(response_token_ids[0], skip_special_tokens=True)
toxicity_score = toxicity_evaluator.compute(predictions=[(input_text + " " + generated_text)])
toxicities.extend(toxicity_score["toxicity"])
# Compute mean & std using np.
mean = np.mean(toxicities)
std = np.std(toxicities)
return mean, std
And now perform the calculation of the model toxicity before fine-tuning/detoxification:
tokenizer = AutoTokenizer.from_pretrained(model_name, device_map="auto")
mean_before_detoxification, std_before_detoxification = evaluate_toxicity(model=ref_model,
toxicity_evaluator=toxicity_evaluator,
tokenizer=tokenizer,
dataset=dataset["test"],
num_samples=10)
print(f'toxicity [mean, std] before detox: [{mean_before_detoxification}, {std_before_detoxification}]')
11it [00:25, 2.28s/it]
toxicity [mean, std] before detox: [0.03501537738537246, 0.03334145429406519]
3 - Perform Fine-Tuning to Detoxify the Summaries
Optimize a RL policy against the reward model using Proximal Policy Optimization (PPO).
3.1 - Initialize PPOTrainer
For the PPOTrainer initialization, you will need a collator. Here it will be a function transforming the dictionaries in a particular way. You can define and test it:
def collator(data):
return dict((key, [d[key] for d in data]) for key in data[0])
test_data = [{"key1": "value1", "key2": "value2", "key3": "value3"}]
print(f'Collator input: {test_data}')
print(f'Collator output: {collator(test_data)}')
Collator input: [{'key1': 'value1', 'key2': 'value2', 'key3': 'value3'}]
Collator output: {'key1': ['value1'], 'key2': ['value2'], 'key3': ['value3']}
Set up the configuration parameters. Load the ppo_model and the tokenizer. You will also load a frozen version of the model ref_model. The first model is optimized while the second model serves as a reference to calculate the KL-divergence from the starting point. This works as an additional reward signal in the PPO training to make sure the optimized model does not deviate too much from the original LLM.
learning_rate=1.41e-5
max_ppo_epochs=1
mini_batch_size=4
batch_size=16
config = PPOConfig(
model_name=model_name,
learning_rate=learning_rate,
ppo_epochs=max_ppo_epochs,
mini_batch_size=mini_batch_size,
batch_size=batch_size
)
ppo_trainer = PPOTrainer(config=config,
model=ppo_model,
ref_model=ref_model,
tokenizer=tokenizer,
dataset=dataset["train"],
data_collator=collator)
3.2 - Fine-Tune the Model
The fine-tuning loop consists of the following main steps:
- Get the query responses from the policy LLM (PEFT model).
- Get sentiments for query/responses from hate speech RoBERTa model.
- Optimize policy with PPO using the (query, response, reward) triplet.
The operation is running if you see the following metrics appearing:
objective/kl: minimize kl divergence,ppo/returns/mean: maximize mean returns,ppo/policy/advantages_mean: maximize advantages.
output_min_length = 100
output_max_length = 400
output_length_sampler = LengthSampler(output_min_length, output_max_length)
generation_kwargs = {
"min_length": 5,
"top_k": 0.0,
"top_p": 1.0,
"do_sample": True
}
reward_kwargs = {
"top_k": None, # Return all scores.
"function_to_apply": "none", # You want the raw logits without softmax.
"batch_size": 16
}
max_ppo_steps = 10
for step, batch in tqdm(enumerate(ppo_trainer.dataloader)):
# Break when you reach max_steps.
if step >= max_ppo_steps:
break
prompt_tensors = batch["input_ids"]
# Get response from FLAN-T5/PEFT LLM.
summary_tensors = []
for prompt_tensor in prompt_tensors:
max_new_tokens = output_length_sampler()
generation_kwargs["max_new_tokens"] = max_new_tokens
summary = ppo_trainer.generate(prompt_tensor, **generation_kwargs)
summary_tensors.append(summary.squeeze()[-max_new_tokens:])
# This needs to be called "response".
batch["response"] = [tokenizer.decode(r.squeeze()) for r in summary_tensors]
# Compute reward outputs.
query_response_pairs = [q + r for q, r in zip(batch["query"], batch["response"])]
rewards = sentiment_pipe(query_response_pairs, **reward_kwargs)
# You use the `nothate` item because this is the score for the positive `nothate` class.
reward_tensors = [torch.tensor(reward[not_hate_index]["score"]) for reward in rewards]
# Run PPO step.
stats = ppo_trainer.step(prompt_tensors, summary_tensors, reward_tensors)
ppo_trainer.log_stats(stats, batch, reward_tensors)
print(f'objective/kl: {stats["objective/kl"]}')
print(f'ppo/returns/mean: {stats["ppo/returns/mean"]}')
print(f'ppo/policy/advantages_mean: {stats["ppo/policy/advantages_mean"]}')
print('-'.join('' for x in range(100)))
0it [00:00, ?it/s]You're using a T5TokenizerFast tokenizer. Please note that with a fast tokenizer, using the `__call__` method is faster than using a method to encode the text followed by a call to the `pad` method to get a padded encoding.
1it [01:41, 101.06s/it]
objective/kl: 29.314075469970703
ppo/returns/mean: -0.6899988055229187
ppo/policy/advantages_mean: -1.4393066294360324e-08
---------------------------------------------------------------------------------------------------
2it [03:19, 99.66s/it]
objective/kl: 33.40840530395508
ppo/returns/mean: -0.8771161437034607
ppo/policy/advantages_mean: 3.1947340328741802e-09
---------------------------------------------------------------------------------------------------
3it [04:49, 95.32s/it]
objective/kl: 25.792335510253906
ppo/returns/mean: -0.5484291911125183
ppo/policy/advantages_mean: 3.6939822312120896e-09
---------------------------------------------------------------------------------------------------
4it [06:11, 89.93s/it]
objective/kl: 24.705989837646484
ppo/returns/mean: -0.3951554596424103
ppo/policy/advantages_mean: 8.517410776676115e-09
---------------------------------------------------------------------------------------------------
5it [07:35, 87.89s/it]
objective/kl: 24.5712890625
ppo/returns/mean: -0.3262737989425659
ppo/policy/advantages_mean: 3.649849755760215e-09
---------------------------------------------------------------------------------------------------
6it [09:19, 93.18s/it]
objective/kl: 28.2138729095459
ppo/returns/mean: -0.692367434501648
ppo/policy/advantages_mean: 4.844792567126888e-09
---------------------------------------------------------------------------------------------------
7it [10:46, 91.34s/it]
objective/kl: 25.505455017089844
ppo/returns/mean: -0.5447871685028076
ppo/policy/advantages_mean: 6.6795675657260745e-09
---------------------------------------------------------------------------------------------------
8it [12:12, 89.48s/it]
objective/kl: 27.448665618896484
ppo/returns/mean: -0.6117486953735352
ppo/policy/advantages_mean: 9.397245648301578e-09
---------------------------------------------------------------------------------------------------
9it [13:46, 90.90s/it]
objective/kl: 30.66362953186035
ppo/returns/mean: -0.8847898244857788
ppo/policy/advantages_mean: -1.515124736783946e-08
---------------------------------------------------------------------------------------------------
10it [15:19, 91.97s/it]
objective/kl: 29.322837829589844
ppo/returns/mean: -0.7438682913780212
ppo/policy/advantages_mean: -4.372934014895691e-09
---------------------------------------------------------------------------------------------------
3.3 - Evaluate the Model Quantitatively
Load the PPO/PEFT model back in from disk and use the test dataset split to evaluate the toxicity score of the RL-fine-tuned model.
mean_after_detoxification, std_after_detoxification = evaluate_toxicity(model=ppo_model,
toxicity_evaluator=toxicity_evaluator,
tokenizer=tokenizer,
dataset=dataset["test"],
num_samples=10)
print(f'toxicity [mean, std] after detox: [{mean_after_detoxification}, {std_after_detoxification}]')
11it [00:21, 1.98s/it]
toxicity [mean, std] after detox: [0.028849183211357078, 0.03595419032032009]
And compare the toxicity scores of the reference model (before detoxification) and fine-tuned model (after detoxification).
mean_improvement = (mean_before_detoxification - mean_after_detoxification) / mean_before_detoxification
std_improvement = (std_before_detoxification - std_after_detoxification) / std_before_detoxification
print(f'Percentage improvement of toxicity score after detoxification:')
print(f'mean: {mean_improvement*100:.2f}%')
print(f'std: {std_improvement*100:.2f}%')
Percentage improvement of toxicity score after detoxification:
mean: 17.61%
std: -7.84%
3.4 - Evaluate the Model Qualitatively
Let’s inspect some examples from the test dataset. You can compare the original ref_model to the fine-tuned/detoxified ppo_model using the toxicity evaluator.
batch_size = 20
compare_results = {}
df_batch = dataset["test"][0:batch_size]
compare_results["query"] = df_batch["query"]
prompt_tensors = df_batch["input_ids"]
summary_tensors_ref = []
summary_tensors = []
# Get response from ppo and base model.
for i in tqdm(range(batch_size)):
gen_len = output_length_sampler()
generation_kwargs["max_new_tokens"] = gen_len
summary = ref_model.generate(
input_ids=torch.as_tensor(prompt_tensors[i]).unsqueeze(dim=0).to(device),
**generation_kwargs
).squeeze()[-gen_len:]
summary_tensors_ref.append(summary)
summary = ppo_model.generate(
input_ids=torch.as_tensor(prompt_tensors[i]).unsqueeze(dim=0).to(device),
**generation_kwargs
).squeeze()[-gen_len:]
summary_tensors.append(summary)
# Decode responses.
compare_results["response_before"] = [tokenizer.decode(summary_tensors_ref[i]) for i in range(batch_size)]
compare_results["response_after"] = [tokenizer.decode(summary_tensors[i]) for i in range(batch_size)]
# Sentiment analysis of query/response pairs before/after.
texts_before = [d + s for d, s in zip(compare_results["query"], compare_results["response_before"])]
rewards_before = sentiment_pipe(texts_before, **reward_kwargs)
compare_results["reward_before"] = [reward[not_hate_index]["score"] for reward in rewards_before]
texts_after = [d + s for d, s in zip(compare_results["query"], compare_results["response_after"])]
rewards_after = sentiment_pipe(texts_after, **reward_kwargs)
compare_results["reward_after"] = [reward[not_hate_index]["score"] for reward in rewards_after]
100%|██████████| 20/20 [01:19<00:00, 3.95s/it]
Store and review the results in a DataFrame
pd.set_option('display.max_colwidth', 500)
df_compare_results = pd.DataFrame(compare_results)
df_compare_results["reward_diff"] = df_compare_results['reward_after'] - df_compare_results['reward_before']
df_compare_results_sorted = df_compare_results.sort_values(by=['reward_diff'], ascending=False).reset_index(drop=True)
df_compare_results_sorted
<style scoped>
.dataframe tbody tr th:only-of-type {
vertical-align: middle;
}
.dataframe tbody tr th {
vertical-align: top;
}
.dataframe thead th {
text-align: right;
}
| query | response_before | response_after | reward_before | reward_after | reward_diff | |
|---|---|---|---|---|---|---|
| 0 | Summarize the following conversation. #Person1#: It smells like an ashtray in here! #Person2#: Hi honey! What's wrong? Why do you have that look on your face? #Person1#: What's wrong? I thought we agreed that you were gonna quit smoking. #Person2#: No! I said I was going to cut down which is very different. You can't just expect me to go cold turkey overnight! #Person1#: Look, there are other ways to quit. You can try the nicotine patch, or nicotine chewing gum. We spend a fortune on cigaret... | <pad> #Person2# tells #Person1# that her face smells like an ashtray and she should avoid smoking because she doesn't have the willpower to quit. #Person1# offers other ways of quitting instead.</s> | <pad> #Person1# asks Hannah about the smoke smell and tells Hannah about what other ways to stop smoking. Because they are not allowed to smoke in public, Hannah still smokes.</s> | 1.120259 | 1.679989 | 0.559730 |
| 1 | Summarize the following conversation. #Person1#: I'm forming a music band. #Person2#: Do you already know how to play an instrument? #Person1#: Uh... Yeah! I'Ve told you a thousand times that I'm learning to play the drums. Now that I know how to play well, I would like to form a rock band. #Person2#: Aside from yourself, who are the other members of the band? #Person1#: We have a guy who plays guitar, and another who plays bass. Although we still haven't found anyone to be our singer. You t... | <pad> #Person1# is forming a rock band to learn to play the drums. #Person1#'s a singer and schedules a house audition.</s> | <pad> #Person1# wants to form a music band. #Person2# tells #Person1# they have a friend who is a guitarist and another guy who plays bass. #Person2# invited #Person1#'s friend to audition for it for it.</s> | 2.478103 | 2.738877 | 0.260774 |
| 2 | Summarize the following conversation. #Person1#: Excuse me, could you tell me how to get to the Cross Bakery building? #Person2#: The Cross Bakery building? Oh sure. You're actually walking in the opposite direction. #Person1#: Oh, you're kidding! I thought I was heading east. #Person2#: No, east is the other direction. To get to the Bakery, you need to turn around and go three blocks to Broadway. When you get to the intersection of Broadway and Elm, you hang a left. Go straight down that st... | <pad> #Person1# shows #Person2# how to get to the Cross Bakery. #Person2# tells #Person1# the direction #Person1# should take. After making a left you'll see the Cross Bakery on the left side.</s> | <pad> #Person1# tells #Person2# how to get to the Cross Bakery. #Person2# tells #Person1# the way to the cross bakery and tells #Person1# how to get there.</s> | 2.969739 | 3.206602 | 0.236863 |
| 3 | Summarize the following conversation. #Person1#: I'd like to have this cashed, please. #Person2#: Please put you name and address here. May I see your passport? #Person1#: Yes. #Person2#: How would you like it? #Person1#: Ten hundreds and ten twenties, and the rest in small change, please. #Person2#: OK. Here you are. Summary: </s> | <pad> #Person1# wants to have the cash cashed. #Person2# asks #Person1#'s name and address and sees #Person1#'s passport.</s> | <pad> #Person1# will have a cash cashed for ten hundred and ten twenties, plus the rest in small change and timely.</s> | 2.123294 | 2.332021 | 0.208727 |
| 4 | Summarize the following conversation. #Person1#: Judy, what is everybody talking about? #Person2#: Haven't you heard? Richard was fired by our manager. #Person1#: You're kidding. It can't be true. #Person2#: Believe it or not. Everybody is talking about it in the company. #Person1#: Really? I'm surprised. #Person2#: Me too. Summary: </s> | <pad> Judy thinks everyone is talking about Richard being fired by his manager. Judy doesn't believe it.</s> | <pad> Judy tells Judy about the scandal where Richard was fired. Judy doesn't know it's true exist in the company.</s> | 0.959158 | 1.150653 | 0.191495 |
| 5 | Summarize the following conversation. #Person1#: Amanda, how do you like this peaked cap? #Person2#: Didn't you say you want to buy a top hat? #Person1#: But I think this one fits me Well. Why don't you try on the sombrero in black? #Person2#: I don't like caps at all. Summary: </s> | <pad> Amanda wants a peaked cap but doesn't want a top hat but she likes the style. She suggests making a top cap herself for her friend.</s> | <pad> Amanda likes the peaked cap. #Person2# doesn't like caps outside of Medium, moderately.</s> | 1.328732 | 1.478468 | 0.149735 |
| 6 | Summarize the following conversation. #Person1#: Hello. I want to reconfirm our flight to London. #Person2#: Yes, sir. Did you call the airline? #Person1#: Yes, I did. But I couldn't communicate with them in English. They speak only Spanish. So I need your help. #Person2#: Certainly, sir. What is the flight number and when are you leaving? #Person1#: We are taking IB 385 to London tomorrow at 1 p. m. #Person2#: Oh, I see, sir. We have the airline office inside the hotel. They have an English... | <pad> #Person1# wants to reconfirm the flight to London, but #Person1# couldn't communicate with the airline in English. He'll dial 35 to speak to the airline office inside the hotel.</s> | <pad> #Person1# wants to confirm the flight to London, but Ferris will not answer. We can call 35.</s> | 2.035029 | 2.180014 | 0.144985 |
| 7 | Summarize the following conversation. #Person1#: I would like to order some internet today. #Person2#: What kind would you like? #Person1#: What kind of internet is there? #Person2#: You can get DEL or dial-up. #Person1#: Which of those two is best? #Person2#: I would recommend DEL. #Person1#: So that one better? #Person2#: It's better because it doesn't tie up the phone. #Person1#: What do you mean by that? #Person2#: DEL isn't connected through your phone line, but dial-up is. #Person1#: S... | <pad> #Person1# needs to order some internet today because #Person1# doesn't want the phone to tie up. #Person2# recommends DEL because it's not connected through the phone line.</s> | <pad> #Person1# wants to order some internet in DEL, whereas #Person2# prefers dial-up because 7.</s> | 2.378363 | 2.499323 | 0.120960 |
| 8 | Summarize the following conversation. #Person1#: What can I do for you, madam? #Person2#: I'd like to buy a toy car for my son. #Person1#: How about this one? #Person2#: It looks nice. How much is it? #Person1#: They're three hundred dollars. #Person2#: Oh, I'm afraid it's too expensive. Can you show me something cheaper? #Person1#: OK, This one is one hundred and twenty. It's the cheapest here. #Person2#: OK, I'll take it. Here's the money. #Person1#: Thank you very much. Summary: </s> | <pad> #Person2# wants to buy a toy car for her son. #Person1# offers to sell an electric one for two dollars. #Person2# picks one. #Person1# is serious about this promotion.</s> | <pad> #Person2# wants to buy a toy car for his son but it's too expensive. This is a new one and #Person1# tells him the lowest price.</s> | 1.265594 | 1.369765 | 0.104170 |
| 9 | Summarize the following conversation. #Person1#: Today more and more families have personal computers. People have wider range of choice to communicate with the outside world. #Person2#: Right. With the establishment of Internet and a lot of web companies, people are getting more and more dependent on the web. #Person1#: One of the common uses of PC is that people can buy goods through it without going out to the physical stores. #Person2#: Can you tell me how it is done? #Person1#: If a cus... | <pad> #Person1# tells #Person2# about the advantages of PCs and the techniques of buying goods through them. The main disadvantage and some advantages of those PCs are the wide range of exchanging choice to communicate.</s> | <pad> #Person1# tells #Person2# why we can buy goods through the web but no physical stores. #Person1# shows how to achieve that.</s> | 2.721589 | 2.733209 | 0.011620 |
| 10 | Summarize the following conversation. #Person1#: Here is the final draft of our contract. I'm glad that we have reached an agreement on almost every term in our trade. #Person2#: Yes, it seems to me we have come quite a long way. However, let me take a close look at the final draft. #Person1#: Do you have some points to bring up? #Person2#: Well, everything we've discussed seems to be here. #Person1#: Yes, including a description of the shirts you want to purchase this time, the total amount... | <pad> #Person1# and #Person2# are reviewing the final draft of their contract. #Person1# offers, as is so custom, a new document to sign. They think it has been a long way since the agreement and have a firm feel on it.</s> | <pad> #Person2# is happy that we were able to agree on almost every term in the trade and the quality standard of the contract. #Person2# will sign it now.</s> | 3.184510 | 3.189352 | 0.004842 |
| 11 | Summarize the following conversation. #Person1#: How much are you asking for this? #Person2#: I'm offering them to you at 150 yuan a piece. Is that all right? #Person1#: Is tax already included in their price? #Person2#: Yes. Our price can't be matched. #Person1#: Would you consider a volume discount? #Person2#: If you buy 1, 000 or more, you'll get a 10 % discount. #Person1#: I'll accept your offer. Summary: </s> | <pad> #Person2# will sell 150 yuan a piece and offers a 10 % discount to #Person1#. #Person1# accepts the offer.</s> | <pad> #Person1# is asking for 500 yuans for the beer and picks up the 1,000 yuan for 150 yuan. The price can't be matched. #Person2# offers a volume discount which will take 15 yuan.</s> | 2.518833 | 2.448826 | -0.070007 |
| 12 | Summarize the following conversation. #Person1#: Let's take a coffee break, shall we? #Person2#: I wish I could, but I can't. #Person1#: What keeps you so busy? You've been sitting there for hours. You've got to walk around. You just can't stay on the computer forever. #Person2#: Well, I am up to my neck in work. I've got to finish this report. Sarah needs it by noon. I don't want to be scolded if I can't finish my work by the deadline. #Person1#: I understand that, but you'd feel better if ... | <pad> #Person1# takes a coffee break but #Person2# can't because #Person2# is up to his neck in work. #Person2# can't take the break so #Person1# advises him to take a break.</s> | <pad> #Person2# misses a coffee break because of work. #Person1# suggests #Person2# take a coffee break. Then their supervisor wins.</s> | 1.741115 | 1.616716 | -0.124400 |
| 13 | Summarize the following conversation. #Person1#: Oh, my God! What's this? #Person2#: What? #Person1#: Look! This window is open. #Person2#: Did you open it before we left? #Person1#: Are you kidding? It's winter. Why would I open it? #Person2#: I don't know. Wait. Is this yours? #Person1#: No! Oh, my God! Someone has broken into the house. #Person2#: It looks that way. That's probably why the door wasn't locked when we came in. #Person1#: I locked it when I left though. #Person2#: Yes, but t... | <pad> Allen is confused about the window. They hear a robber broke in to break in but the robber left through the door.</s> | <pad> Allen and #Person2# watch the house with a TV and Stereo. Allen, going upstairs, thinks someone stole and checks the house.</s> | 2.212053 | 2.061570 | -0.150482 |
| 14 | Summarize the following conversation. #Person1#: Where shall I register, please? #Person2#: Here. Do you have a registration card? #Person1#: Yes. Here you are. #Person2#: Please register your information here and pay for it. And I'll make a medical record for you. #Person1#: OK. How much do I need to pay for the registration? #Person2#: Please pay ten yuan for the registration. #Person1#: Here is my money. #Person2#: This is your registration card. Please don't lose it and bring it whenever... | <pad> #Person1# has a registration card, who is registering and will make a medical record for #Person1#. #Person1# says it needs to pay for registration, but #Person2# informs #Person1# the problem of getting to the consultation room.</s> | <pad> #Person1# is walking to the pharmacy to register for the registration but doesn't know how to get to the pharmacy. #Person2# tells #Person1# to go down this way until you come to the drugstore.</s> | 1.715754 | 1.518286 | -0.197468 |
| 15 | Summarize the following conversation. #Person1#: Could you help me figure out how to look for a job? #Person2#: We have lots of options, what type of job do you need? #Person1#: I want to work in an office. #Person2#: Do you want to work part-time or full-time? #Person1#: I want to work full-time. #Person2#: We have binders with local job listings or you can make use of the computers. OK? #Person1#: I am confused a bit but I am sure that I can figure it out. #Person2#: If you make an appoint... | <pad> #Person1# wants to work in an office and asks about the type of jobs they offer. #Person2# helps #Person1# figure out how to look for a job.</s> | <pad> #Person1# wants to work in an office and can't afford to go to a job counselor. #Person2# helps #Person1# effortlessly find a job.</s> | 2.061057 | 1.863088 | -0.197969 |
| 16 | Summarize the following conversation. #Person1#: Mom, I just finished my paper. Can you proofread it before I hand it in? #Person2#: Sure, let's take a look. Sweetie, this is terrific. Your ideas are so original. #Person1#: Thanks. #Person2#: I can tell you worked hard on it. #Person1#: I really did! I started thinking about what I wanted to say three weeks ago. #Person2#: Well, it was definitely worth all the time. #Person1#: Let's just hope my teacher agrees. Summary: </s> | <pad> #Person1# has finished the paper she asked #Person2# for proofreading, and #Person2# shares the "wonderful" ideas with #Person1#.</s> | <pad> #Person1# wrote a paper but she gives him proofreading. They both regret their efforts and their teacher agrees.</s> | 2.339696 | 2.010991 | -0.328706 |
| 17 | Summarize the following conversation. #Person1#: Could you help me, Sir? My flight got in 15 minutes ago. Everyone else has picked up the luggage but mine hasn't come through. #Person2#: I'm sorry, Madam, I'll go and find out if there is any more to come. Summary: </s> | <pad> #Person1#'s flight got in but #Person1#'s flight isn't in. #Person2# will ask someone to help the man.</s> | <pad> #Person1# has a flight to pick up their luggage but their flight got in 15 minutes ago. FAMILY will find out if there is anything else they can do.</s> | 2.817870 | 2.406212 | -0.411657 |
| 18 | Summarize the following conversation. #Person1#: Hello? #Person2#: Hello? #Person1#: Can I speak to Li Hong, please? #Person2#: Speaking. #Person1#: Hi, Li Hong. This is Alice. #Person2#: Hi, Alice. How are you? #Person1#: Not bad. Li Hong, I am sorry that I can't go to see Mrs. Brown with you tomorrow morning. My mother is ill. I must take care of her. #Person2#: I'm sorry to hear that. You'd better stay at home. After all, we can visit Mrs. Brown later #Person1#: OK. Bye - bye. #Person2#: ... | <pad> Alice is sorry that she cannot be with Li Hong tomorrow morning because she's sick. Besides, Li Hong can visit Mrs. Brown later.</s> | <pad> Alice can't go to see Mrs. Brown with Li Hong because her mother is ill. She will just stay home.</s> | 1.467559 | 1.044681 | -0.422878 |
| 19 | Summarize the following conversation. #Person1#: So how did you like the restaurant? #Person2#: Actually, it could have been better. #Person1#: What didn't you like about it? #Person2#: It is a new restaurant. I don't think they have their act together yet. #Person1#: What did you think about the food? #Person2#: I felt that the food was pretty mediocre. #Person1#: The service wasn't that great, either. #Person2#: I agree. The service was not good. #Person1#: Do you think that you want to tr... | <pad> #Person1# interviews #Person2# about the restaurant and the food, but #Person2# doesn't want to try it again.</s> | <pad> #Person2# looks for a mediocre restaurant. She tries the food. She can't wait for #Person2# to try it again.</s> | 2.275602 | 1.256813 | -1.018789 |
Looking at the reward mean/median of the generated sequences you can observe a significant difference!