Improving Probability-based Prompt Selection Through Unified Evaluation and Analysis

Official code for the paper Improving Probability-based Prompt Selection Through Unified Evaluation and Analysis accepted at TACL 2024.

You can read the summary of our paper in this Twitter Thread.

Use the following to cite our paper:

jsx @article{yang2024improving, title={Improving Probability-based Prompt Selection Through Unified Evaluation and Analysis}, author={Yang, Sohee and Kim, Jonghyeon and Jang, Joel and Ye, Seonghyeon and Lee, Hyunji and Seo, Minjoon}, journal={TACL}, year={2024} }

This repository provides a set of tools for easy utilization and unified evaluation of different probability-based prompt selection methods.

If you're interested in reproducing the experimental results mentioned in the paper, please refer to the Reproduction section.

Features

1. Extraction of the language model's output probability necessary for prompt selection score calculation.
2. Probability-based prompt selection methods.
3. Calibration methods.
4. Task-wise prompt selection and Instance-wise prompt selection.
5. Custom prompt addition that conforms to the template format of promptsource.

Installation

Installation from the source. Python's virtual or Conda environments are recommended.

bash git clone git@github.com:soheeyang/unified-prompt-selection.git cd unified-prompt-selection pip install -r requirements.txt

Python version 3.9+ is required.

Overview

LLMs predict the essential $p(y|x,t)$ through inference with given prompt candidates and datasets to calculate prompt selection scores. In the subsequent prompt selection process, the extracted $p(y|x,t)$ is loaded to calculate prompt selection scores. Based on these scores, a prompt is chosen, and the selection result is returned. OTR(One-Token Response) Converter is used when calculating $p(y|x,t)$ by utilizing only the first token logits.

Quick Start

You can quickly explore the core functionalities by referring to the following notebook files.

• Prompt Selection

  

• Adding Custom Prompt Selection Method

  

• Adding Custom Prompt

  

LLM Inference & Prompt Selection

By running run_prompt_selection.py, you can extract $p(y|x,t)$ and select a prompt.

The extracted $p(y|x,t)$ through LLM inference is stored in the './extraction/results' directory. When running run_prompt_selection.py with the same configuration, it utilizes the previously saved extraction results without additional inference. For detailed information about inference, please refer to the $p(y|x,t)$ Extraction section.

bash python run_prompt_selection.py Running the command as shown above will execute Prompt Selection according to the predefined default arguments.

You can also execute various combinations by adding -m or --multirun as follows: bash python run_prompt_selection.py -m \ method=MI,MI_G,MI_L,MI_GL,GE,GE_M,LE,MDL,MDL_M,ZLP,ZPM,ZMV,PPL,PPL_L \ calibration=cbm-softmax,cbm-mean,cc-softmax,cc-mean,pmi-softmax,pmi-mean \ decoder=opt-1.3b,opt-2.7b,opt-6.7b,opt-30b,opt-66b,gpt-neo-1.3b,gpt-neo-2.7b,gpt-j-6b,gpt2-xl,bloom-3b \ dataset=sst2,ag_news,cb,imdb,newspop,rte,sst5,tweet_emotion,tweet_irony,piqa,copa,hellaswag,story_cloze \ prompt=base_prompts,v12_prompts,v2_prompts,fewshot_prompt \ first_token=false,true \ sum_log_prob=false,true \ fewshot=null,'1,2,4' \ filter=false,true \ unbalance=false,true

We used hydra to manage complex configurations. You can check the configurations in ./conf, and besides specifying arguments on the command line, you can modify the arguments by editing the ./conf/config.yaml file.

After executing python run_prompt_selection.py, you can verify the prompt selection result through the CLI. The result of executing the above command is as follows:

```bash [2023-10-30 02:38:35,738][main][INFO] - -------------------------------------------------------------------------------------------------------------------------------------------------------------------------- Evaluation results using predictions from selected prompt.

* X
   Accuracy:   0.5906
   F1 score:   0.5001
   Prediction: [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] ...
   Target:     [1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1, 1, 0, 1] ...
   Correct:    [1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1, 1, 0, 1] ...

* A
   Accuracy:   0.9186
   F1 score:   0.9183
   Prediction: [1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1] ...
   Target:     [1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1, 1, 0, 1] ...
   Correct:    [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1] ...

* P
   Accuracy:   0.5906
   F1 score:   0.5001
   Prediction: [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1] ...
   Target:     [1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1, 1, 0, 1] ...
   Correct:    [1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 0, 1, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1, 1, 0, 1] ...

* PA
   Accuracy:   0.9186
   F1 score:   0.9183
   Prediction: [1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1] ...
   Target:     [1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 0, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1, 1, 0, 1] ...
   Correct:    [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1] ...

Note that some datasets are missing label values, 
so check the target results.

The predictions of the selected prompt were saved in the following file.

'./results/dataset=glue_sst2_validation__decoder=facebook--opt-2.7b__prompt=base_prompts__first_token=False__sum_log_prob=False__num_samples=1000__seed=42__fewshot=None__do_eval=True/method=MI__all_tokens=True__one_hot=False__select_for_each_x=False__cali_type=cbm__cali_norm_type=softmax__filter=False__unbalance=False.json'

--------------------------------------------------------------------------------------------------------------------------------------------------------------------------

```

You can obtain four different results depending on whether calibration is applied:

• X: without applying any calibration

• A: applying calibration only for Answer selection

• P: applying calibration only for Prompt selection

• PA: applying calibration for both Prompt selection and Answer selection.

The prompt selection result is saved in the ‘./results/dataset=""__decoder=""__prompt=""__first_token=""__sum_log_prob=""__num_samples=""__seed=""__fewshot=""__do_eval=""/method=""__using_all_tokens=""__one_hot=""__select_for_each_x=""__filter=""__unbalance="".json’ file. The format of the result file is as follows:

Notes: - When applying prompt selection to a dataset without ground truth labels, prompt selection is possible, but evaluation results cannot be verified. Before applying prompt selection to a dataset, check whether ground truth labels are available or not or check whether the evaluation results for datasets without ground truth labels are based on the Target of the prompt selection results.

Probability-based Prompt Selection Method

The following probability-based prompt selection methods are available: 'MI', 'GE', 'LE', 'MDL', 'ZLP', 'ZPM', 'ZMV', and 'PPL'.

To use a specific prompt selection method, pass the desired method to method. You can find detailed descriptions of each method in section 2.2 Existing Approaches of the paper.

Variants created by Prompt Selection Methods

The following methods are variants that modify the score calculation formula of existing Probability-based prompt selection methods: 'MIG', 'MIL,' 'MIGL', 'GEM', 'MDLM', and 'PPLL'.

You can check the arguments specific to these probability-based prompt selection methods in the ./conf/method directory.

If a method name is followed by '_L', it means that select_for_each_x is set to 'True', and instance-wise prompt selection is performed. The methods that support instance-wise prompt selection are 'MDL', 'MI', and 'PPL'.

If a method name is followed by '_G', it means that one_hot is set to 'True', and one-hot $p(y|x,t)$ is used for GE calculation.

Adding Custom Prompt Selection Method

If you want to add a new prompt selection method, refer to the ./method/methods.py module. Check other methods in this module, and create a new method according to the types and dimensions of input and output values.

||Task-wise Method|Instance-wise Method| |:---------:|:-------:|:------------:| |Input Variable|template_prob|tensor_dict_prob| |Input Variable Type|torch.Tensor|torch.Tensor| |Input Variable Size|[X, Y]|[T, X, Y]| |Output Value Type|torch.Tensor|Tuple[List[float], List[int]]| |Output Value Size|[](Scalar)|([X], [X])|

• T : The number of prompts
• X : The number of instances
• Y : The number of answer choices

The Instance-wise Method has two lists as output values. The first list should contain the results of calculated instance-wise prompt selection scores using the added method, and the second list should contain the indices of the selected prompts for each instance.

The newly added method is utilized in the get_*_wise_ps_scores function in the ./method/score.py module.

python methodFuncMap = { 'MI': get_mi_g if one_hot else get_mi, 'GE': get_ge if one_hot else get_ge_m, 'LE': get_le, 'MDL': get_le, 'PPL': get_ppl, 'ZLP': get_zlp, 'ZPM': get_zpm, 'ZMV': get_zmv, }

When adding a method, follow the format of the method in the methodFuncMap inside the get_*_wise_ps_result function. Additionally, add a yaml file representing the new method to ./conf/method directory. Refer to the yaml file format of other methods in ./conf/method

Calibration Method

Three calibration methods are available, allowing you to observe the changes in prompt selection results before and after applying each calibration.

You can check the arguments related to calibration in the ./conf/calibration directory.

You can choose one of the following calibration methods by specifying 'cbm', 'cc', or 'pmi' to cali_type. The default value is 'cbm'.

• cbm: Calibration By Marginalization, an enhanced calibration method proposed in this work that demonstrates effective results in prompt selection compared to existing methods.
• cc: Contextual Calibration, a calibration method proposed in Calibrate before use: Improving Few-Shot performance of language models.
• pmi: Domain Conditional PMI, a calibration method proposed in Surface form competition: Why the highest probability answer isn’t always right.

You can specify the normalization criteria for $\tilde{q}(y|x,t)$ using cali_norm_type. The options are 'softmax' or 'mean', and the default value is 'softmax'.

Dynamic Dataset

Dynamic datasets are datasets where answer choices are not fixed and vary for each instance. The experiment includes COPA, PIQA, StoryCloze, and Hellaswag as dynamic datasets. With such datasets, you can adjust the label distribution differently and observe the changes in prompt selection results based on the label distribution.

If you set unbalance to 'True', you can adjust the label distribution of COPA, PIQA, StoryCloze, and Hellaswag to be unbalanced at a ratio of 100:0 and observe the prompt selection results.

Filter

In Zero-Label Prompt Selection, a low-quality prompt filtering algorithm using k-means clustering and $p(y|x,t)$ was proposed. By setting filter to 'True', prompts can be filtered using the zero-label prompt selection (ZPS) filtering algorithm.

$p(y|x,t)$ Extraction

To perform probability-based prompt selection, you need to extract the output probability of the language model, denoted as $p(y|x,t)$, where $x,t$ represents the instantiated prompt.

If desired, it is also possible to perform LLM inference excluding prompt selection and extract $p(y|x,t)$ using the run_inference.py script.

bash python run_inference.py \ decoder=opt-2.7b \ dataset=sst2 \ prompt=base_prompts \ first_token=false \ sum_log_prob=false \ num_samples=1000 \ seed=42 \ fewshot=null \ do_eval=true \ mixed_precision=no

The arguments that affect extraction can be found in ./conf/config.yaml.

To calculate $p(y|x,t)$ using only the first token of $y$, set the first_token to 'True'. If set to 'False', all tokens of $y$ are used.

When calculating $p(y|x,t)$ using all tokens over $y$, you have the option to calculate either the mean log-probability or the sum of log-probability for all tokens over $y$. To extract the sum of log-probability for all tokens over $y$, set sum_log_prob to 'True'.

To adjust the dataset sample size for prompt selection, use the num_samples.

For reproducibility, you can set the random seed by using the seed.

If you want to select one of the fewshot prompts with a different number and order of examples for a single prompt template, you can use fewshot.

Through the do_eval, you can choose whether to record evaluation results ('Accuracy', 'Macro F1') for each prompt during the inference stage for extraction. If set to 'False,' prompt selection can still be performed even if there is no ground truth label in the dataset. However, you will only be able to verify what the final selected prompt is and its predictions.

Notes: - The results of $p(y|x,t)$ extraction are saved under the ./extraction/results/dataset=""__decoder=""__prompt=""__first_token=""__sum_log_prob=""__num_samples=""__seed=""__fewshot=""__do_eval="" directory. - If results have already been extracted with the same arguments (i.e., the directory with the same name exists under ./extraction/results), the existing results will be provided, and additional extraction will not be performed. - If a directory with the same name exists but the names of the selected prompt templates have changed in the settings, additional extraction will only be carried out for prompts with names that were not previously extracted. - $p(y|x,t)$ extraction results from each prompt will be saved as JSON files under the directory. The output file has the following format:

Decoder

If you want to specify a particular model, you can pass the desired model name to the decoder. The models used in the experiment can be found in ./conf/decoder.

You can also add new models by adding .yaml files under ./conf/decoder. You can refer to existing files to configure the arguments.

Pass the desired model name to model_name_or_path. Adding models is limited to the decoder models supported by the Hugging Face transformers library.

❗ Verify the pad_token_id of the tokenizer to be used and pass the verified pad_token_id to ignore_index.

To apply model parallelism, use the parallelize.

You can adjust the batch size by using the per_device_eval_batch_size.

Dataset

If you want to specify a particular dataset, you can pass the desired dataset name to dataset. The datasets used in the experiment can be found in ./conf/dataset.

You can also add new datasets by adding .yaml files under ./conf/dataset. You can refer to existing files to configure the arguments. To manage and utilize prompts, we rely on promptsource. Therefore, adding new datasets is limited to classification datasets existing in promptsource.

dataset_name corresponds to the path argument of the load_dataset method in the Hugging Face datasets library.

dataset_config_name corresponds to the name argument.

split corresponds to the split argument.

DATASET_KWARGS are automatically determined based on the configured dataset_name, dataset_config_name, and split.

DATASET_INFO is required for OTR (=One Token Response, first_token=True). - num_classes should be set to the number of label categories. - label should be set to the column name containing ground truth labels for the dataset. If it doesn't exist, you can set it to null. In such cases, don't forget to set the do_eval option to 'False' as well. - is_dynamic should be set to True if the dataset you want to add is a dynamic task; otherwise, set it to False. For an explanation of dynamic tasks, refer to Dynamic Dataset. - choices should be set to values corresponding to the answer_choices associated with the dataset's prompt templates.

TEMPLATE_INFO is required for Custom Prompt Addition. Set text_format to the column name containing text in the appropriate format and jinja_suffix to the column name containing ground truth labels in the format specified by jinja.

Note:

• If you want to apply a dataset that is not present in promptsource, you will need a templates.yaml file that adheres to the format used by promptsource. We recommend creating this file by referring to examples in the promptsource documentation and templates.yaml files for other datasets. Also, please note that if there are no ground truth labels, you can create the jinja section without adding jinja_suffix.

Prompt

You can find the list of prompts used in the experiment in the .yaml files under conf/prompt.

The prompts you want to use can be configured through prompt. Create a .yaml file under conf/prompt and pass the name of the created file to prompt.

prompt_config_name affects the directory name where the results of $p(y|x,t)$ extraction are stored ('prompt=&{prompt_config_name}').

By entering the names of prompt templates as a list in template_names, you can select prompts from those templates.

If you want to add new prompts, refer to Custom Prompt Addition.

Adding Custom Prompt

To manage and utilize prompts, we rely on promptsource. For more information on adding prompts, refer to promptsource.

You can add a prompt by running add_prompt.py.

Here is an example of adding a prompt for the 'ag_news' dataset:

bash python add_prompt.py dataset=ag_news

Instruction

When running python add_prompt.py dataset=ag_news, follow the instructions below:

``` Instruction example: "{{ text }}" is about

Enter the instruction for the new prompt, using the example format provided above: ```

The placeholder '{{ ... }}' represents the location where the instance of the dataset will be inserted. The ... refers to the column name that contains the instances for each dataset.

``` Instruction example: "{{ text }}" is about

Enter the instruction for the new prompt, using the example format provided above: {{ text }} this is testaddprompt.

Please verify that the entered instruction is correct: {{ text }} this is testaddprompt.

If the instruction is correct, press 'y'; otherwise, enter another key. If you choose another key, you will be able to edit the instruction: y ```

Follow the provided example and input the desired instruction accordingly.

Notes: -

Answer Choices

``` Answer_choices example: politics ||| sports ||| business ||| science

Enter the answer_choices for the new prompt, using the example format provided above: ```

After entering the instruction, input the answer_choices.

``` Answer_choices example: politics ||| sports ||| business ||| science

Enter the answer_choices for the new prompt, using the example format provided above: politics ||| sports ||| business ||| science

Please verify that the entered answer_choices are correct: politics ||| sports ||| business ||| science

If the answerchoices are correct, press 'y'; otherwise, enter another key. If you choose another key, you will be able to edit the answerchoices: y ```

Enter the desired answer_choices according to the provided example.

Notes: -

• If the number of answerchoices is different from the number of labels, you will need to enter the answerchoices again.

``` Answer_choices example: politics ||| sports ||| business ||| science

Enter the answer_choices for the new prompt, using the example format provided above: politics ||| sports

The number of answerchoices must be 4. Please enter the answerchoices again. ```