Skip to content

Latest commit

 

History

History
344 lines (272 loc) · 14.1 KB

README.md

File metadata and controls

344 lines (272 loc) · 14.1 KB
Raw

Reformatted Alignment

This is the official repository for Reformatted Alignment.

Run-Ze Fan, Xuefeng Li, Haoyang Zou, Junlong Li, Shwai He, Ethan Chern, Jiewen Hu, Pengfei Liu

News

  • Sep 2024: Our paper has been accepted by EMNLP 2024 Findings! 🎉
  • Feb 2024: We release the preprint paper on Arxiv, ReAlign data, and other useful resources in developing them (tasks description, hand-written format, tasks classifier, training data, and NQ dataset for factuality evaluation).

Table of contents

Introduction

We explores elevating the quality of existing instruction data to better align with human values, introducing a simple and effective approach named ReAlign (Reformatted Alignment), which reformats the responses of instruction data into a format that better aligns with pre-established criteria and the collated evidence. This approach minimizes human annotation, hallucination, and the difficulty in scaling, remaining orthogonal to existing alignment techniques. Experimentally, ReAlign significantly boosts the general alignment ability, math reasoning, factuality, and readability of the LLMs.

Encouragingly, without introducing any additional data or advanced training techniques, and merely by reformatting the response, LLaMA-2-13B's mathematical reasoning ability on GSM8K can be improved from 46.77% to 56.63% in accuracy. Additionally, a mere 5% of ReAlign data yields a 67% boost in general alignment ability measured by the Alpaca dataset. This work highlights the need for further research into the science and mechanistic interpretability of LLMs.

The underlying philosophy of ReAlign is to re-coordinate the roles of humans and LLMs in the alignment process, leveraging their complementary strengths -- humans articulate their preferences, and LLMs, in turn, reconstruct instructions based on their generative power (e.g., instruction-following ability), without directly using distilled LLM knowledge. Through this collaborative synergy, we expect the generated instruction data to be not only more contextually precise but also more closely aligned with human preferences.

The accuracy of the GSM8K test set for LLaMA-2-13B and Mistral-7B models fine-tuned on the training set of GSM8K and MATH with and without ReAlign. (a): Training and testing on GSM8K. (b): Training on MATH and testing on GSM8K (Out-of-Distribution Setting).
An overview of our ReAlign including three steps. KILT denotes Knowledge Intensive Language Tasks.

The ReAlign process unfolds in three main steps.

The first step involves criteria definition, where humans define their preferences (e.g., the preferred format of responses) in various scenarios in the form of natural language. In this paper, we meticulously define criteria for 46 distinct scenarios.

The second step, retrieval augmentation, broadens the knowledge base for knowledge-intensive tasks like open-domain QA and fact verification. This is achieved by incorporating additional information, thereby improving the factuality and informativeness of responses.

The final step, reformatting, aims to re-align the responses with the pre-established criteria and the collated evidence, guaranteeing outputs that are both structured and substantiated.

ReAlign realigns the original response with the pre-defined criteria to be a better format.
An example of the response from the original model and the response from the ReAlign Model

Quick Start

Setup

We use python 3.10 in this project. You are encouraged to create a virtual environment through conda.

Then, we have to install all the libraries listed in requirements.txt. Note that you may choose an appropriate version of torch according to your CUDA version (we write torch>=2.0.1+cu118 in this file).

pip install -r requirements.txt

Pipeline

  • get your OpenAI API key from here. This is used for reformatting.
  • get your Serper API key from here. This is only used for retrieval with Google Search.

Step 1: Task Classification

Download the task classifier from huggingface hub:

Model Name HF Checkpoints Size License
Task Classifier 🤗 GAIR/ReAlign-Task-Classifier 13B Llama 2

Then, by using the following prompt, the task classifier can identify which task a query belongs to:

PROMPT_INPUT_FOR_TASK_CLS: str = '''
You will receive a user's query. Additionally, you are given some pre-defined tasks below: 

[Existing tasks start]
question_generation
story_generation
poem_generation
email_generation
data_generation
advice_giving
recommendations
how_to_generation
planning
instructional_rewriting
language_polishing
paraphrasing
text_correction
code_correction
code_simplification
information_extraction
keywords_extraction
table_extraction
title_generation
text_summarization
note_summarization
explain_code
explain_answer
text_to_text_translation
text_to_code_translation
code_to_code_translation
code_to_text_translation
open_qa
closed_qa
fill_in_the_blank
fact_verification
math_puzzles
language_learning_questions
natural_language_learning_tutor
exam_problem_solving_tutor
ml_ai_language_model_tutor
general_classification
ordering
sentiment_analysis
code_language_classification
language_classification
topic_classification
value_judgement
rejecting
roleplay
default
[Existing tasks end]

You objective is to choose the most appropriate task that can reflect the high-level intention of this query. You should first clearly give out your choice. Your choice should exactly match one of the task names provided above, without any modification. Do not include the task description in your choice.

Your output should be just the task name.

User's query is below:
[User's query start]
{input}
[User's query end]

Task name:

'''

Here is an example:

from vllm import LLM, SamplingParams
import torch

num_gpus = torch.cuda.device_count()
model_name_or_dir = "GAIR/ReAlign-Task-Classifier" # or the local directory to store the downloaded model
llm = LLM(model=model_name_or_dir, tensor_parallel_size=num_gpus)

query = "Give three tips for staying healthy."
input_ = PROMPT_INPUT_FOR_TASK_CLS.format(input=query)

sampling_params = SamplingParams(temperature=0.0, top_p=1.0, max_tokens=50)
outputs = llm.generate(input_, sampling_params)
task = output[0].outputs[0].text

print(task) # should be `advice_giving`.
# If the classified result is not in task list, set it as `default`.

Step 2: Prepare your dataset

Convert your dataset into the following format with json type, same as the ReAlign datasets.

Here is an example:

[
    {
        "id": 0,
        "items": [
            {
                # question
                "from": "human",
                "value": "Give three tips for staying healthy.",
                "category": "advice_giving"
            },
            {
                # response
                "from": "gpt",
                "value": "1.Eat a balanced diet and make sure to include plenty of fruits and vegetables. \n2. Exercise regularly to keep your body active and strong. \n3. Get enough sleep and maintain a consistent sleep schedule."
            }
        ]
    }
]

Step 3: Retrieval with Google Search

Set your Serper API key:

export SERPER_API_KEY=...

Run the following script:

python retrieval.py \
    --input_data_path dataset.json \
    --output_path dataset_retrieval.json \
    --batch_size 10

The output file:

dataset_retrieval.jsonis added the original retrieval results.

dataset_retrieval_clean_evidence.jsonis added the cleaned retrieval results. This is used for ReAlign.

Step 4: Reformat

Set your OpenAI API key:

export OPENAI_API_KEY=...

Run the following script:

python reformat.py \
    --input_data_path dataset_retrieval_clean_evidence.json \
    --output_directory reformat_results \
    --tokenizer_path meta-llama/Llama-2-7b-chat-hf \ # or the local directory to store the downloaded tokenizer
    --dataset_batch_id 0 \ # the first file (it's in 0 - 9) of ten files
    --dataset_batch_num 10 \ # the total number of the file
    --openai_key <OPENAI_API_KEY> \
    --top_k 2 \ # output 2 reformatted response for each response
    --model gpt-3.5-turbo-1106 \
    --temperature 0.3 \
    --top_p 1 \
    --target_length 4096

Note that we are using process parallel to speedup, which means that we are going to run dataset_batch_num processes at the same time for reformatting, and each process will need to specify the dataset_batch_id manually.

For example:

If you set dataset_batch_num as 10, it means the datasets will be split into 10 subdataset (10x acceleration). You should run the script 10 times at the same time, each time specifying dataset_batch_id as 0 through 9.

Then, you can get dataset_batch_num files in the directory output_directory.

Run the following script to merge these files into one final datasets:

python parallel_data_merge.py \
    --input_data_path dataset_retrieval_clean_evidence.json \ # the <input_data_path> in reformat script
    --output_directory reformat_results \ # the <output_directory> in reformat script
    --final_output_path dataset_reformat.json

Finally, you can get the final reformatted datasets.

Step 5: Post Filtering

You can combine the filtering rules in rewrite_data_selection.py or customize the filtering rules.

Run the following script to filter the reformatted dataset:

python rewrite_data_selection.py \
    --input_original_data_path dataset_retrieval_clean_evidence.json \ # the dataset path before reformatting
    --input_rewrite_data_path dataset_reformat.json \ # the reformatted dataset path
    --output_path realign_dataset.json # the final dataset path after filtering

Now, you can get the final realign dataset realign_dataset.json.

ReAlign Dataset

We reformat five datasets based Open-Platypus, Alpaca, No Robots, GSM8K, and MATH:

ReAlign Open-Platypus: datasets/realign_OpenPlatypus.json

ReAlign Alpaca: datasets/realign_alpaca.json

ReAlign No Robots: datasets/realign_no_robots.json

ReAlign GSM8K: datasets/realign_gsm8k.json

ReAlign MATH: datasets/realign_math.json

The datasets also can be loaded on 🤗Hugging Face:

Dataset Name Hugging Face Link Size
ReAlign Open-Platypus 🤗 GAIR/ReAlign-Open-Platypus 25K
ReAlign Alpaca 🤗 GAIR/ReAlign-Alpaca 52K
ReAlign No Robots 🤗 GAIR/ReAlign-No-Robots 10K
ReAlign GSM8K 🤗 GAIR/ReAlign-GSM8K 7.4K
ReAlign MATH 🤗 GAIR/ReAlign-MATH 6.5K

Other Resources

Tasks Description and Formats

The tasks description and predefined formats can be found in code/constant.py.

The Data for Task Classifier

The training data for the task classifier is in datasets/classification/task_classifier_train_dataset.json.

The test data is in datasets/classification/task_classifier_test_dataset.json.

The format is as follows:

{
        "instruction": "Create a story about a dog that finds a magical portal.",
        "category": "story_generation"
}

Factuality Evaluation

We randomly sample 100 cases from NQ Dataset for factuality evaluation, which can be found in datasets/nq.

The ground truth is in datasets/nq/nq_factuality_100.json.

The format is as follows:

{
        "items": [
            {
                "from": "human",
                "value": "when did the democratic party change its name?"
            },
            {
                "from": "gpt",
                "value": "the 1830s"
            }
        ],
        "id": 0
}

Citation

Please cite the paper if the resource in this repo or the paper is helpful to you.

@article{fan2024reformatted,
      title={Reformatted Alignment}, 
      author={Fan, Run-Ze and Li, Xuefeng and Zou, Haoyang and Li, Junlong and He, Shwai and Chern, Ethan and Hu, Jiewen and Liu, Pengfei},
      year={2024},
      journal={arXiv preprint arXiv:2402.12219},
      url={https://arxiv.org/abs/2402.12219}
}

Acknowledgements

We thank the GAIR members for reviewing our paper and giving valuable feedback. We appreciate the authors in OpenChat for providing the training codebase and the helpfulness.