Giter VIP home page Giter VIP logo

voidism / dola Goto Github PK

View Code? Open in Web Editor NEW
386.0 3.0 47.0 7.64 MB

Official implementation for the paper "DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models"

Home Page: https://arxiv.org/abs/2309.03883

Python 91.16% Makefile 0.01% Dockerfile 0.04% Jsonnet 0.01% Shell 0.14% MDX 7.84% Jupyter Notebook 0.42% C++ 0.03% Cuda 0.34% Cython 0.01% C 0.01%
factuality hallucinations large-language-models

dola's Introduction

DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models

License: MIT Arxiv Hugging Face Transformers Tweet GitHub Stars

Open In Colab

Code for the paper "DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models"

Paper: https://arxiv.org/abs/2309.03883
Authors: Yung-Sung Chuang $^\dagger$, Yujia Xie $^\ddagger$, Hongyin Luo $^\dagger$, Yoon Kim $^\dagger$, James Glass $^\dagger$, Pengcheng He $^\ddagger$
$^\dagger$ Massachusetts Institute of Technology, $^\ddagger$ Microsoft

Overview

DoLa

Despite their impressive capabilities, large language models (LLMs) are prone to hallucinations, i.e., generating content that deviates from facts seen during pretraining. We propose a simple decoding strategy for reducing hallucinations with pretrained LLMs that does not require conditioning on retrieved external knowledge nor additional fine-tuning. Our approach obtains the next-token distribution by contrasting the differences in logits obtained from projecting the later layers versus earlier layers to the vocabulary space, exploiting the fact that factual knowledge in an LLMs has generally been shown to be localized to particular transformer layers. We find that this Decoding by contrasting Layers (DoLA) approach is able to better surface factual knowledge and reduce the generation of incorrect facts. DoLA consistently improves the truthfulness across multiple choices tasks and open-ended generation tasks, for example improving performance of LLaMA family models on TruthfulQA by 12-17% absolute points, demonstrating its potential in making LLMs reliably generate truthful facts.

Setup

pip install -e transformers-4.28.1
pip install datasets
pip install accelerate
pip install openai # -> only for truthfulqa and gpt4_eval

Experiments

Arguments

Argument Example Description
--model-name huggyllama/llama-7b Specifies the model you want to use, currently we only support LLaMA-v1.
--data-path /path/to/dataset Path to the dataset file or folder.
--output-path output-path.json Where to store the output results.
--num-gpus 1 Number of GPUs to use, 1/2/4/8 for 7B/13B/30B/65B model sizes respectively.
--max_gpu_memory 27 Maximum GPU memory size (in GiB) to allocate. Default: 27 (for 32G V100).

Understanding --early-exit-layers

The --early-exit-layers argument takes a string containing a sequence of layer numbers separated by commas, with no spaces in between. By specifying different number of layers, we make the model decode at different modes.

Number of Layers Specified Example (str) Description of Decoding Mode
1 -1 Naive decoding from the final layer output.
2 16,32 DoLa-static decoding with the second specified layer (i.e. 32) as the mature_layer and first specified layer (i.e. 16) as premature_layer.
>2 0,2,4,6,8,10,12,14,32 DoLa decoding with the last specified layer (i.e. 32) as the mature_layer and all the preceding layers (i.e. 0,2,4,6,8,10,12,14) as candidate_premature_layers.

FACTOR (Multiple Choices)

Please download the data file wiki_factor.csv from https://github.com/AI21Labs/factor

Baseline

python factor_eval.py --model-name huggyllama/llama-7b --data-path /path/to/wiki_factor.csv --output-path output-path.json --num-gpus 1
python factor_eval.py --model-name huggyllama/llama-13b --data-path /path/to/wiki_factor.csv --output-path output-path.json --num-gpus 2
python factor_eval.py --model-name huggyllama/llama-30b --data-path /path/to/wiki_factor.csv --output-path output-path.json --num-gpus 4
python factor_eval.py --model-name huggyllama/llama-65b --data-path /path/to/wiki_factor.csv --output-path output-path.json --num-gpus 8

DoLa

python factor_eval.py --model-name huggyllama/llama-7b --early-exit-layers 0,2,4,6,8,10,12,14,32 --data-path /path/to/wiki_factor.csv --output-path output-path.json --num-gpus 1
python factor_eval.py --model-name huggyllama/llama-13b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,40 --data-path /path/to/wiki_factor.csv --output-path output-path.json --num-gpus 2
python factor_eval.py --model-name huggyllama/llama-30b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,60 --data-path /path/to/wiki_factor.csv --output-path output-path.json --num-gpus 4
python factor_eval.py --model-name huggyllama/llama-65b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,80 --data-path /path/to/wiki_factor.csv --output-path output-path.json --num-gpus 8

TruthfulQA (Multiple Choices)

The --data-path should be a folder contains TruthfulQA.csv. If file not exists, it will be downloaded automatcially.

Baseline

python tfqa_mc_eval.py --model-name huggyllama/llama-7b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 1
python tfqa_mc_eval.py --model-name huggyllama/llama-13b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 2
python tfqa_mc_eval.py --model-name huggyllama/llama-30b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 4
python tfqa_mc_eval.py --model-name huggyllama/llama-65b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 8

DoLa

python tfqa_mc_eval.py --model-name huggyllama/llama-7b --early-exit-layers 16,18,20,22,24,26,28,30,32 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 1
python tfqa_mc_eval.py --model-name huggyllama/llama-13b --early-exit-layers 20,22,24,26,28,30,32,34,36,38,40 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 2
python tfqa_mc_eval.py --model-name huggyllama/llama-30b --early-exit-layers 40,42,44,46,48,50,52,54,56,58,60 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 4
python tfqa_mc_eval.py --model-name huggyllama/llama-65b --early-exit-layers 60,62,64,66,68,70,72,74,76,78,80 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 8

TruthfulQA

To evaluate the open-ended generation result of TruthfulQA, we need to finetune two GPT-3 curie models through OpenAI API:

openai api fine_tunes.create -t finetune_truth.jsonl -m curie --n_epochs 5 --batch_size 21 --learning_rate_multiplier 0.1
openai api fine_tunes.create -t finetune_info.jsonl -m curie --n_epochs 5 --batch_size 21 --learning_rate_multiplier 0.1

After finetuning, we can obtain the finetuned model names by openai api fine_tunes.list | grep fine_tuned_model.

Create a config file gpt3.config.json like this:

{"gpt_info": "curie:ft-xxxxxxxxxx",
"gpt_truth": "curie:ft-xxxxxxxxxx",
"api_key": "xxxxxxx"}

Add the argument --do-rating --gpt3-config gpt3.config.json for GPT-3 evaluation.

Baseline

python tfqa_eval.py --model-name huggyllama/llama-7b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 1 --do-rating --gpt3-config /path/to/gpt3.config.json
python tfqa_eval.py --model-name huggyllama/llama-13b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 2 --do-rating --gpt3-config /path/to/gpt3.config.json
python tfqa_eval.py --model-name huggyllama/llama-30b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 4 --do-rating --gpt3-config /path/to/gpt3.config.json
python tfqa_eval.py --model-name huggyllama/llama-65b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 8 --do-rating --gpt3-config /path/to/gpt3.config.json

DoLa

python tfqa_eval.py --model-name huggyllama/llama-7b --early-exit-layers 16,18,20,22,24,26,28,30,32 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 1 --do-rating --gpt3-config /path/to/gpt3.config.json
python tfqa_eval.py --model-name huggyllama/llama-13b --early-exit-layers 20,22,24,26,28,30,32,34,36,38,40 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 2 --do-rating --gpt3-config /path/to/gpt3.config.json
python tfqa_eval.py --model-name huggyllama/llama-30b --early-exit-layers 40,42,44,46,48,50,52,54,56,58,60 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 4 --do-rating --gpt3-config /path/to/gpt3.config.json
python tfqa_eval.py --model-name huggyllama/llama-65b --early-exit-layers 60,62,64,66,68,70,72,74,76,78,80 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 8 --do-rating --gpt3-config /path/to/gpt3.config.json

GSM8K

We use a random sampled subset of GSM8K training set to serve as the validation set of both StrategyQA and GSM8K. The file can be downloaded here.

The --data-path argument should be either:

  • A folder that contains gsm8k_test.jsonl, otherwise the files will be downloaded automatically to the folder you specified.
  • (Only for GSM8K) The path to gsm8k-train-sub.jsonl which can be downloaded by the link above.

Baseline

python gsm8k_eval.py --model-name huggyllama/llama-7b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 1
python gsm8k_eval.py --model-name huggyllama/llama-13b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 2
python gsm8k_eval.py --model-name huggyllama/llama-30b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 4
python gsm8k_eval.py --model-name huggyllama/llama-65b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 8

DoLa

python gsm8k_eval.py --model-name huggyllama/llama-7b --early-exit-layers 0,2,4,6,8,10,12,14,32 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 1
python gsm8k_eval.py --model-name huggyllama/llama-13b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,40 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 2
python gsm8k_eval.py --model-name huggyllama/llama-30b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,60 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 4
python gsm8k_eval.py --model-name huggyllama/llama-65b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,80 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 8

StrategyQA

The --data-path argument should be a folder that contains strategyqa_train.json, otherwise the files will be downloaded automatically to the folder you specified.

Baseline

python strqa_eval.py --model-name huggyllama/llama-7b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 1
python strqa_eval.py --model-name huggyllama/llama-13b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 2
python strqa_eval.py --model-name huggyllama/llama-30b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 4
python strqa_eval.py --model-name huggyllama/llama-65b --data-path /path/to/data/folder --output-path output-path.json --num-gpus 8

DoLa

python strqa_eval.py --model-name huggyllama/llama-7b --early-exit-layers 0,2,4,6,8,10,12,14,32 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 1
python strqa_eval.py --model-name huggyllama/llama-13b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,40 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 2
python strqa_eval.py --model-name huggyllama/llama-30b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,60 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 4
python strqa_eval.py --model-name huggyllama/llama-65b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,80 --data-path /path/to/data/folder --output-path output-path.json --num-gpus 8

GPT-4 Evaluation (Vicuna QA Benchmark)

In GPT-4 evaluation, we need the question file from FastChat. In the following commands, we assume the path to your FastChat repo is $fastchat.

Baseline

python gpt4_judge_eval.py --model-name huggyllama/llama-7b --model-id llama-7b-baseline --question-file $fastchat/eval/table/question.jsonl --answer-file output-answer.jsonl --num-gpus 1
python gpt4_judge_eval.py --model-name huggyllama/llama-13b --model-id llama-13b-baseline --question-file $fastchat/eval/table/question.jsonl --answer-file output-answer.jsonl --num-gpus 2
python gpt4_judge_eval.py --model-name huggyllama/llama-30b --model-id llama-30b-baseline --question-file $fastchat/eval/table/question.jsonl --answer-file output-answer.jsonl --num-gpus 4
python gpt4_judge_eval.py --model-name huggyllama/llama-65b --model-id llama-65b-baseline --question-file $fastchat/eval/table/question.jsonl --answer-file output-answer.jsonl --num-gpus 8

DoLa

python gpt4_judge_eval.py --model-name huggyllama/llama-7b --early-exit-layers 0,2,4,6,8,10,12,14,32 --model-id llama-7b-dola --question-file $fastchat/eval/table/question.jsonl --answer-file output-answer.jsonl --num-gpus 1
python gpt4_judge_eval.py --model-name huggyllama/llama-13b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,40 --model-id llama-13b-dola --question-file $fastchat/eval/table/question.jsonl --answer-file output-answer.jsonl --num-gpus 2
python gpt4_judge_eval.py --model-name huggyllama/llama-30b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,60 --model-id llama-30b-dola --question-file $fastchat/eval/table/question.jsonl --answer-file output-answer.jsonl --num-gpus 4
python gpt4_judge_eval.py --model-name huggyllama/llama-65b --early-exit-layers 0,2,4,6,8,10,12,14,16,18,80 --model-id llama-65b-dola --question-file $fastchat/eval/table/question.jsonl --answer-file output-answer.jsonl --num-gpus 8

After running the above commands to generate the model responses, we need OpenAI API key to pairwise compare the responses from different decoding results.

python $fastchat/eval/eval_gpt_review.py -q $fastchat/eval/table/question.jsonl -a output-answer-1.jsonl output-answer-2.jsonl -p $fastchat/eval/table/prompt.jsonl -r $fastchat/eval/table/reviewer.jsonl -o output-review-path.jsonl -k openai_api_key

For more details of GPT-4 evaluation, please check vicuna-blog-eval.

Reference Repositories

Citation

DOI

Please cite our paper if it's helpful to your work!

@article{chuang2023dola,
  title={DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models},
  author={Chuang, Yung-Sung and Xie, Yujia and Luo, Hongyin and Kim, Yoon and Glass, James and He, Pengcheng},
  journal={arXiv preprint arXiv:2309.03883},
  year={2023},
}

dola's People

Contributors

hillzhang1999 avatar voidism avatar

Stargazers

avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar avatar

Watchers

avatar avatar avatar

Forkers

kemolo zhangnn520 techthiyanes ai-natural-language-processing-lab stjordanis eltociear jmanhype cizhenshi kttian wangxidong06 iie-ycx knowledgehacker matheusalb codeaudit shobhit-agarwal hsliu-initial trantrikien239 pendulum445 darcstar-solutions-tech hillzhang1999 daoyuan14 ecoli-hit lliutianc refraintc githubarry rampart-aios dbtjr1103 nguyenvuthientrang zencyyoung opensorceycw occasionallynlp hanzhou032 xkpengl itshuey itsmemala young-j-park matteonulli newton-9560 nkwbtb kagaii maxleaf3824 liam0949 fakerno1

dola's Issues

Some questions about the idea of the paper

I'm very interested in your work, but I'm confused about some of the ideas in the paper. In this paper, it is believed that the output of shallow layer pays more attention to grammatical coherence, and LLM will inject more knowledge with the deepening of layers. So you attenuate the illusion of the output by suppressing the distribution of the shallow output. But while I admit that shallow output does lack the necessary knowledge. But this does not mean that the shallow distribution is the inverse of the true distribution. That is, words with small probability of shallow output not must be good, and words with large probability not must be bad. Therefore, I am skeptical of your theoretical interpretation of this contrast method.

How to install the environment if I don't have root permission?

Hello, I'm following your work and trying to run your code. When I install the environment in my anaconda env with pip install -e transformers-4.28.1 I meet the following problem:

error: can't create or remove files in install directory

The following error occurred while trying to add or remove files in the
installation directory:

    [Errno 13] Permission denied: '/usr/local/lib/python3.10/dist-packages/test-easy-install-3876739.write-test'

The installation directory you specified (via --install-dir, --prefix, or
the distutils default setting) was:

    /usr/local/lib/python3.10/dist-packages/

Perhaps your account does not have write access to this directory?  If the
installation directory is a system-owned directory, you may need to sign in
as the administrator or "root" account.  If you do not have administrative
access to this machine, you may wish to choose a different installation
directory, preferably one that is listed in your PYTHONPATH environment
variable.

Can you provide full requirement.txt or provide some help? Thank you!

concerns about your experiments of TruthfulQA-MC

I am very excited with your fantastic work, while there also exisits a little concern about your experiments.

Following your instructions, I can obtain $(MC1, MC2, MC3) = (32.07, 63.78, 32.06)$ for TruthfulQA-MC, which are almost the same as the results you reported in your paper. However, I notice that you set post_softmax=False during evalution. It conflcts with the equation (4) in your paper.

image

I have tried to set post_softmax=True, and the scores drops to $(31.95, 52.21, 28.15)$ but still better than vanilla.

I am concerned that it is not supported to use the logits directly without using a softmax to convert logits to probabilities. I would appreciate it if you have any insights to explain such a confusion.

https://github.com/voidism/DoLa/blob/dc88907406f9744f748f3c779f2353efd5bdc824/tfqa_mc_eval.py#L302C318-L302C336

questions about code

In dola.py line112 get_relative_top_filter()
sorted_logits, sorted_indices = torch.sort(scores_normalized, descending=True)
min_thresh = sorted_logits[..., min_tokens_to_keep-1]
probs_max = torch.max(scores_normalized, dim=-1).values
probs_thresh = probs_max + np.log(relative_top)
In my concern:
is that min_thresh = probs_max ?
and np.log(relative_top) < 0
then probs_thresh must lesser than min_thresh
so whats the meanings of min_thresh ?
Thx

The way to calculate the log_probs

Hi,
As the output of the model in each token's position represents the possibilities of next token, should the calculation of log_probs be misaligned.
I mean "diff_logits[range(diff_logits.shape[0]-1), continue_ids[1:]].sum().item()"
instead of "log_probs = diff_logits[range(diff_logits.shape[0]), continue_ids].sum().item()".

Figure 2 subtoken "_"

How did you guys get subtoken (the underbar) like _W ole _So y ink a _was ... in figure 2.

Should apply model.norm layer to hidden_states[early_exit_layer] ?

DoLa/transformers-4.28.1/src/transformers/models/llama/modeling_llama.py

Line 703 in dc88907

logits = self.lm_head(outputs.hidden_states[early_exit_layer])

I think you guys should apply model.norm layer to hidden_states[early_exit_layer] . Because only the last hidden_state applied model.norm layer. See

DoLa/transformers-4.28.1/src/transformers/models/llama/modeling_llama.py

Line 594 in dc88907

hidden_states = self.norm(hidden_states)

Replacing .sum() with .mean() in dola.py?

Hi Team, great work on the project!

I noticed something in dola.py line 217 that might need attention. The code is:

log_probs = diff_logits[range(diff_logits.shape[0]), continue_ids].sum().item()

I am wondering if .sum() should be replaced with .mean()? Since false answers and correct answers may have different number of tokens, using .mean() might provide a more fair comparison.

Thanks!

Support for LLaMA-2

Hi, nice work! I would like to know which parts of the code you have modified in transformers-4.28.1, and how can I support LLaMA-2?

Query regarding JS Divergence mean over batches.

Hi!
In dola.py on line 197, we calculate js_divs = js_divs.mean(-1) using which premature layer is selected. Does this mean that premature layer is the same for any tokens/logits present in the batch?
Can you help me out with this please? @voidism

data for gpt-judge training

Hi,

Where can I access the below files for fine-tuning gpt-judge for truthful-qa eval please?
finetune_truth.jsonl
finetune_info.jsonl

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.