Skip to content

Latest commit

 

History

History
374 lines (277 loc) · 18.4 KB

README_swebench.md

File metadata and controls

374 lines (277 loc) · 18.4 KB

Agentless experiments on SWE-Bench

🐈Setup | 🙀Localization | 😼Repair | 😸Patch Validation and Selection | 💰Cost

In this document, we will go through the steps to generate the patches on SWE-bench. Currently, we support both SWE-Bench Lite and SWE-Bench Verified

🐈 Setup

First create the environment

git clone https://github.com/OpenAutoCoder/Agentless.git
cd Agentless

conda create -n agentless python=3.11 
conda activate agentless
pip install -r requirements.txt
export PYTHONPATH=$PYTHONPATH:$(pwd)

Then export your OpenAI API key

export OPENAI_API_KEY={key_here}

Create the folder to save our results

mkdir results # where we will save our results

Tip

We currently support SWE-Bench Lite and SWE-Bench Verified benchmarks, you can use --dataset to select the benchmark (by default it will be SWE-Bench Lite)

For example --dataset=princeton-nlp/SWE-bench_Verified

Tip

Since for each issue in the benchmark (both SWE-Bench Lite and SWE-Bench Verified) we need to checkout the repository and process the files, you might want to save some time by downloading the preprocessed data here: swebench_lite_repo_structure.zip

After downloading, please unzip and export the location as such export PROJECT_FILE_LOC={folder which you saved}

Tip

You can use --target_id to specific a particular bug you want to target.

For example --target_id=django__django-10914

Tip

We use multiple threads (controllable via --num_threads) to speed up the Agentless process

🙀 Localization

In localization, the goal is to find the locations in source code where we need to edit to fix the issues. At a high-level, Agentless uses a 3-stage localization step to first localize to specific files, then to relevant code elements, and finally to fine-grained edit locations. We will now take you through the step-by-step procedure of Agentless in each of the localization steps.

1. Localize to suspicious files

First, we localize to suspicious files. This is done in a multi-step process where we combine both LLM localized files with retrieval files.

Run the following command to generate the LLM-predicted suspicious files:

python agentless/fl/localize.py --file_level \
                                --output_folder results/swe-bench-lite/file_level \
                                --num_threads 10 \
                                --skip_existing

This will save all the LLM-predicted suspicious file locations in results/swe-bench-lite/file_level/loc_outputs.jsonl with the logs saved in results/swe-bench-lite/file_level/localization_logs

We then complement the previous suspicious files with a simple embedding-based retrieval method to identify additional suspicious files.

This is done by first filtering out irrelevant folders by using LLM to produce a list of irrelevant folders that do not need to be retrieved from with the following command:

python agentless/fl/localize.py --file_level \
                                --irrelevant \
                                --output_folder results/swe-bench-lite/file_level_irrelevant \
                                --num_threads 10 \
                                --skip_existing

This will save the identified irrelevant folders in results/swe-bench-lite/file_level_irrelevant/loc_outputs.jsonl with the logs saved in results/swe-bench-lite/file_level_irrelevant/localization_logs

We then perform the retrieval (note the embedding is done with OpenAI text-embedding-3-small model) by passing in the irrelevant folders and running the following command:

python agentless/fl/retrieve.py --index_type simple \
                                --filter_type given_files \
                                --filter_file results/swe-bench-lite/file_level_irrelevant/loc_outputs.jsonl \
                                --output_folder results/swe-bench-lite/retrievel_embedding \
                                --persist_dir embedding/swe-bench_simple \
                                --num_threads 10

This will save the retrieved files in results/swe-bench-lite/retrievel_embedding/retrieve_locs.jsonl with the logs saved in results/swe-bench-lite/retrievel_embedding/retrieval_logs

Finally we merge the LLM-predicted suspicious file locations with the embedding-based retrieved files to obtain a final list of relevant files:

python agentless/fl/combine.py  --retrieval_loc_file results/swe-bench-lite/retrievel_embedding/retrieve_locs.jsonl \
                                --model_loc_file results/swe-bench-lite/file_level/loc_outputs.jsonl \
                                --top_n 3 \
                                --output_folder results/swe-bench-lite/file_level_combined

results/swe-bench-lite/file_level_combined/combined_locs.jsonl contains the final list of suspicious files identified by Agentless.

2. localize to related elements

Next, we move on to localizing the related elements.

Run the following command to provide the suspicious files from the first stage as input:

python agentless/fl/localize.py --related_level \
                                --output_folder results/swe-bench-lite/related_elements \
                                --top_n 3 \
                                --compress_assign \
                                --compress \
                                --start_file results/swe-bench-lite/file_level_combined/combined_locs.jsonl \
                                --num_threads 10 \
                                --skip_existing

This will save the related elements in results/swe-bench-lite/related_elements/loc_outputs.jsonl with the logs saved in results/swe-bench-lite/related_elements/localization_logs

3. localize to edit locations

Finally, using the related elements, we then localize to the edit locations. This is done via sampling to obtain multiple different sets of edit locations:

python agentless/fl/localize.py --fine_grain_line_level \
                                --output_folder results/swe-bench-lite/edit_location_samples \
                                --top_n 3 \
                                --compress \
                                --temperature 0.8 \
                                --num_samples 4 \
                                --start_file results/swe-bench-lite/related_elements/loc_outputs.jsonl \
                                --num_threads 10 \
                                --skip_existing

This will save the edit locations in results/swe-bench-lite/edit_location_samples/loc_outputs.jsonl with the logs saved in results/swe-bench-lite/edit_location_samples/localization_logs

⏬ Structure of `loc_outputs.jsonl` :: click to expand ::
  • instance_id: task ID of the issue
  • found_files: list of files localized by the model
  • additional_artifact_loc_file: raw output of the model during file-level localization
  • file_traj: trajectory of the model during file-level localization (e.g., # of tokens)
  • found_related_locs: dict of relevant code elements localized by the model
  • additional_artifact_loc_related: raw output of the model during relevant-code-level localization
  • related_loc_traj: trajectory of the model during relevant-code-level localization
  • found_edit_locs: dict of edit locations localized by the model
  • additional_artifact_loc_edit_location: raw output of the model during edit-location-level localization
  • edit_loc_traj: trajectory of the model during edit-location-level localization

Run the following command to separate the individual sets of edit locations:

python agentless/fl/localize.py --merge \
                                --output_folder results/swe-bench-lite/edit_location_individual \
                                --top_n 3 \
                                --num_samples 4 \
                                --start_file results/swe-bench-lite/edit_location_samples/loc_outputs.jsonl

The separate sets of edit locations can be found in results/swe-bench-lite/edit_location_individual. The location files will be named loc_merged_{x}-{x}_outputs.jsonl where x indicates the individual samples. For our experiments on SWE-bench, we will use all 4 sets of edit locations and perform repairs on them individually to generate 4 different repair runs.

😼 Repair

Using the 4 sets of edit locations from before, we now perform repair.

Agentless generates multiple patches per issue (controllable via parameters) and then perform majority voting with patch validation to select the final patch for submission

Run the following command to generate the patches:

python agentless/repair/repair.py --loc_file results/swe-bench-lite/edit_location_individual/loc_merged_0-0_outputs.jsonl \
                                  --output_folder results/swe-bench-lite/repair_sample_1 \
                                  --loc_interval \
                                  --top_n=3 \
                                  --context_window=10 \
                                  --max_samples 10  \
                                  --cot \
                                  --diff_format \
                                  --gen_and_process \
                                  --num_threads 2
Additional Repair Commands 
for i in {1..3}; do
    python agentless/repair/repair.py --loc_file results/swe-bench-lite/edit_location_individual/loc_merged_${i}-${i}_outputs.jsonl \
                                    --output_folder results/swe-bench-lite/repair_sample_$((i+1)) \
                                    --loc_interval \
                                    --top_n=3 \
                                    --context_window=10 \
                                    --max_samples 10  \
                                    --cot \
                                    --diff_format \
                                    --gen_and_process \
                                    --num_threads 2 
done

These commands generate 10 samples each (1 greedy and 9 via temperature sampling) as defined --max_samples 10. The --context_window indicates the amount of code lines before and after each localized edit location we provide to the model for repair. The patches are saved in results/swe-bench-lite/repair_sample_{i}/output.jsonl, which contains the raw output of each sample as well as any trajectory information (e.g., number of tokens). The complete logs are also saved in results/swe-bench-lite/repair_sample_{i}/repair_logs/

The above commands will combine to generate 40 samples in total for each bug.

😸 Patch Validation and Selection

Since Agentless generates multiple candidate patches per issue, we need a way to select a final patch for submission.

To do this, Agentless leverages both regression tests that exist in the codebase as well as generating new reproduction tests that can verify if the patch can solve the original issue.

Regression test selection

We first select a set of regression tests (tests that already exist in the repository and pass on the original codebase) to run.

Run the following command to get a list of passing tests in the original codebase:

python agentless/test/run_regression_tests.py --run_id generate_regression_tests \
                                              --output_file results/swe-bench-lite/passing_tests.jsonl

This will generate a list of passing tests at results/swe-bench-lite/passing_tests.jsonl

Note

We do not use any of provided PASS_TO_PASS field in the SWE-bench benchmark

We select tests from the complete list of tests which can pass in the original repository

Next, we ask the LLM to remove any tests which should not be ran with the following command:

python agentless/test/select_regression_tests.py --passing_tests results/swe-bench-lite/passing_tests.jsonl \
                                                 --output_folder results/swe-bench-lite/select_regression

This will produce a list of final regression tests in results/swe-bench-lite/select_regression/output.jsonl with the logs at results/swe-bench-lite/select_regression/select_test_logs

We can run this on all the patches generate, repeated for each repair run (i.e., by changing folder):

folder=results/swe-bench-lite/repair_sample_1
for num in {0..9..1}; do
    run_id_prefix=$(basename $folder); 
    python agentless/test/run_regression_tests.py --regression_tests results/swe-bench-lite/select_regression/output.jsonl \
                                                  --predictions_path="${folder}/output_${num}_processed.jsonl" \
                                                  --run_id="${run_id_prefix}_regression_${num}" --num_workers 10;
done

This will output the regression test results in the same folder as the repair results. results/swe-bench-lite/repair_sample_1/output_{i}_regression_test_results.jsonl contains the regression test results for each patch number (i).

Note

We also perform post-processing to generate the complete git-diff patch for each repair sample.

You can find the individual patch in results/repair/output_{i}_processed.jsonl where i is the sample number.

Reproduction test generation

In addition to the regression tests, Agentless also generates a reproduction test that attempt to check if the patch can solve the original issue.

Similar to patch generation, Agentless also generates multiple samples of reproduction tests, and then perform selection:

python agentless/test/generate_reproduction_tests.py --max_samples 40 \
                                                     --output_folder results/swe-bench-lite/reproduction_test_samples \
                                                     --num_threads 10

This will generate 40 samples (1 greedy + 39 temperature sampling) per issue. The generated reproduction tests can be found in results/swe-bench-lite/reproduction_test_samples/output.jsonl. The corresponding logs can be found in results/swe-bench-lite/reproduction_test_samples/generating_test_logs/.

Now we will execute each of these generated tests on the original repository to see if they can reproduce the original issue.

for st in {0..36..4}; do   en=$((st + 3));   
        echo "Processing ${st} to ${en}";   
        for num in $(seq $st $en); do     
            echo "Processing ${num}";     
            python agentless/test/run_reproduction_tests.py --run_id="reproduction_test_generation_filter_sample_${num}" \
                                                            --test_jsonl="results/swe-bench-lite/reproduction_test_samples/output_${num}_processed_reproduction_test.jsonl" \
                                                            --num_workers 6 \
                                                            --testing;
done & done

Warning

In the above command we execute multiple SWE-bench evaluations in parallel, please ensure that your machine is able to handle that

If not, you may want to reduce the amount of parallelization

This produces verification results for each tests in the same folder: results/swe-bench-lite/reproduction_test_samples/

We then perform majority voting to select one reproduction test per issue:

python agentless/test/generate_reproduction_tests.py --max_samples 40 \
                                                     --output_folder results/swe-bench-lite/reproduction_test_samples \
                                                     --output_file reproduction_tests.jsonl \
                                                     --select

This will generate the reproduction test file at: results/swe-bench-lite/reproduction_test_samples/reproduction_tests.jsonl

Finally, we evaluate the generated patches on the selected reproduction test. Similar to regression test execution, this is repeated for each repair run (i.e., by changing folder):

folder=results/swe-bench-lite/repair_sample_1
for num in {0..9..1}; do
    run_id_prefix=$(basename $folder); 
    python agentless/test/run_reproduction_tests.py --test_jsonl results/swe-bench-lite/reproduction_test_samples/reproduction_tests.jsonl \
                                                    --predictions_path="${folder}/output_${num}_processed.jsonl" \
                                                    --run_id="${run_id_prefix}_reproduction_${num}" --num_workers 10;
done

This will output the reproduction test results in the same folder as the repair results. results/swe-bench-lite/repair_sample_1/output_{i}_reproduction_test_results.jsonl contains the reproduction test results for each patch number (i).

Reranking and patch selection

Finally, using the regression and reproduction test results, Agentless performs reranking to select the final patch for submission.

Run the following command (--regression indicates we are using regression tests for selection --reproduction indicates we are using the reproduction tests for selection)

python agentless/repair/rerank.py --patch_folder results/swe-bench-lite/repair_sample_1/,results/swe-bench-lite/repair_sample_2/,results/swe-bench-lite/repair_sample_3/,results/swe-bench-lite/repair_sample_4/ \
                                  --num_samples 40 \
                                  --deduplicate \
                                  --regression \
                                  --reproduction

This command will produced the all_preds.jsonl that contains the final selected patch for each instance_id which you can then directly use your favorite way of testing SWE-bench for evaluation!

💰 Cost

To measure the cost of running Agentless, we have provided helpful utilities.

For each of the output.jsonl files produced for each of the steps (including substeps), run the following command:

python dev/util/cost.py --output_file example_step/output.jsonl

This will output the dollar cost as well as the number of tokens. --embedding_cost can be used to compute the cost of the embedding step.