README.md

---
license: apache-2.0
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: source
    dtype: string
  - name: file_name
    dtype: string
  - name: cwe
    sequence: string
  splits:
  - name: train
    num_bytes: 1015823
    num_examples: 113
  download_size: 405079
  dataset_size: 1015823
---

# SOTA fine-tuning by OpenAI

OpenAI used the [synth-vuln-fixes](https://huggingface.co/datasets/patched-codes/synth-vuln-fixes) and fine-tuned
a new version of gpt-4o is now the SOTA on this benchmark. More details and code is available from their [repo.](https://github.com/openai/build-hours/tree/main/5-4o_fine_tuning)

![graph](openai-static-analysis-eval.png)

More details on the benchmark are available in our [blog](https://www.patched.codes/blog/the-static-analysis-evaluation-benchmark-measuring-llm-performance-in-fixing-software-vulnerabilities).

# New Version of Static Analysis Eval (Aug 20, 2024)

We have created a new version of the benchmark with instances that are harder than the previous one. There has been a lot of progress in models
over the last year as a result the previous version of the benchmark was saturated. The methodology is the same, we have also released the 
dataset generation script which scans the top 100 Python projects to generate the instances. You can see it [here](_script_for_gen.py). 
The same [eval script](_script_for_eval.py) works as before. You do not need to login to Semgrep anymore as we 
only use their OSS rules for this version of the benchmark.

The highest score a model can get on this benchmark is 100%, you can see the oracle run logs [here](oracle-0-shot_semgrep_1.85.0_20240820_174931.log).

# New Evaluation

| Model | Score | Logs |
|:-----:|:-----:|:----:|
| gpt-4o-mini | 52.21 | [link](gpt-4o-mini-0-shot_semgrep_1.85.0_20240820_201236.log)|
| gpt-4o-mini + 3-shot prompt | 53.10 | [link](gpt-4o-mini-3-shot_semgrep_1.85.0_20240820_213814.log)| 
| gpt-4o-mini + rag (embedding & reranking) | 58.41 | [link](gpt-4o-mini-3-shot-sim_semgrep_1.85.0_20240821_023541.log) |
| gpt-4o-mini + fine-tuned with [synth-vuln-fixes](https://huggingface.co/datasets/patched-codes/synth-vuln-fixes) | 53.98 | [link](ft_gpt-4o-mini-2024-07-18_patched_patched_9yhVV00P-0-shot_semgrep_1.85.0_20240821_082958.log) |


| Model | Score | Logs |
|:-----:|:-----:|:----:|
| gpt-4o | 53.10 | [link](gpt-4o-0-shot_semgrep_1.85.0_20240820_210136.log)|
| gpt-4o + 3-shot prompt | 53.98 | [link](gpt-4o-3-shot_semgrep_1.85.0_20240820_215534.log)| 
| gpt-4o + rag (embedding & reranking) | 56.64 | [link](gpt-4o-3-shot-sim_semgrep_1.85.0_20240821_025455.log) |
| gpt-4o + fine-tuned with [synth-vuln-fixes](https://huggingface.co/datasets/patched-codes/synth-vuln-fixes) | 61.06 | [link](ft_gpt-4o-2024-08-06_patched_patched_9yhZp9nn-0-shot_semgrep_1.85.0_20240821_084452.log) |


## Mixture of Agents (MOA)

We also benchmarked gpt-4o with [Patched MOA](https://arxiv.org/abs/2407.18521). This demostrates that an inference optimization 
technique like MOA can improve performance without fine-tuning.

| Model | Score | Logs |
|:-----:|:-----:|:----:|
| moa-gpt-4o | 53.98 | [link](moa-gpt-4o-2024-08-06-0-shot_semgrep_1.85.0_20240824_032808.log)|
| moa-gpt-4o + 3-shot prompt | 60.18 | [link](moa-gpt-4o-2024-08-06-3-shot_semgrep_1.85.0_20240824_035842.log)| 
| moa-gpt-4o + rag (embedding & reranking) | 61.06 | [link](moa-gpt-4o-2024-08-06-3-shot-sim_semgrep_1.85.0_20240824_043304.log) |

# Static Analysis Eval Benchmark

A dataset of 76 Python programs taken from real Python open source projects (top 100 on GitHub), 
where each program is a file that has exactly 1 vulnerability as detected by a particular static analyzer (Semgrep).

You can run the `_script_for_eval.py` script to check the results.

```
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python _script_for_eval.py
```

For all supported options, run with `--help`:

```
usage: _script_for_eval.py  [-h] [--model MODEL] [--cache] [--n_shot N_SHOT] [--use_similarity] [--oracle]

Run Static Analysis Evaluation

options:
  -h, --help        show this help message and exit
  --model MODEL     OpenAI model to use
  --cache           Enable caching of results
  --n_shot N_SHOT   Number of examples to use for few-shot learning
  --use_similarity  Use similarity for fetching dataset examples
  --oracle          Run in oracle mode (assume all vulnerabilities are fixed)
```

We need to use the logged in version of Semgrep to get access to more rules for vulnerability detection. So, make sure you login before running the eval script.

```
% semgrep login
API token already exists in /Users/user/.semgrep/settings.yml. To login with a different token logout use `semgrep logout`
```

After the run, the script will also create a log file which captures the stats for the run and the files that were fixed.
You can see an example [here](gpt-4o-mini_semgrep_1.85.0_20240818_215254.log).
Due to the recent versions of Semgrep not detecting a few of the samples in the dataset as vulnerable anymore, the maximum score
possible on the benchmark is 77.63%. You can see the oracle run log [here](oracle-0-shot_semgrep_1.85.0_20240819_022711.log).

## Evaluation
We did some detailed evaluations recently (19/08/2024):

| Model | Score | Logs |
|:-----:|:-----:|:----:|
| gpt-4o-mini | 67.11 | [link](gpt-4o-mini_semgrep_1.85.0_20240818_215254.log)|
| gpt-4o-mini + 3-shot prompt | 71.05 | [link](gpt-4o-mini-3-shot_semgrep_1.85.0_20240818_234709.log)| 
| gpt-4o-mini + rag (embedding & reranking) | 72.37 | [link](gpt-4o-mini-1-shot-sim_semgrep_1.85.0_20240819_013810.log) |
| gpt-4o-mini + fine-tuned with [synth-vuln-fixes](https://huggingface.co/datasets/patched-codes/synth-vuln-fixes) | 77.63 | [link](ft_gpt-4o-mini-2024-07-18_patched_patched_9uUpKXcm_semgrep_1.85.0_20240818_220158.log) |


| Model | Score | Logs |
|:-----:|:-----:|:----:|
| gpt-4o | 68.42 | [link](gpt-4o-0-shot_semgrep_1.85.0_20240819_015355.log)|
| gpt-4o + 3-shot prompt | 77.63 | [link](gpt-4o-3-shot_semgrep_1.85.0_20240819_020525.log)| 
| gpt-4o + rag (embedding & reranking) | 77.63 | [link](gpt-4o-1-shot-sim_semgrep_1.85.0_20240819_023323.log) |
| gpt-4o + fine-tuned with [synth-vuln-fixes](https://huggingface.co/datasets/patched-codes/synth-vuln-fixes) | 77.63 | [link](ft_gpt-4o-2024-05-13_patched_patched-4o_9xp8XOM9-0-shot_semgrep_1.85.0_20240819_075205.log) |

# Leaderboard

The top models on the leaderboard are all fine-tuned using the same dataset that we released called [synth vuln fixes](https://huggingface.co/datasets/patched-codes/synth-vuln-fixes).
You can read about our experience with fine-tuning them on our [blog](https://www.patched.codes/blog/a-comparative-study-of-fine-tuning-gpt-4o-mini-gemini-flash-1-5-and-llama-3-1-8b).
You can also explore the leaderboard with this [interactive visualization](https://claude.site/artifacts/5656c16d-9751-407c-9631-a3526c259354).
![Visualization of the leaderboard](visualization.png)

|           Model           | StaticAnalysisEval (%) |  Time (mins)  | Price (USD) |
|:-------------------------:|:----------------------:|:-------------:|:-----------:|
|   gpt-4o-mini-fine-tuned   |        77.63           |     21:0      |    0.21     |
| gemini-1.5-flash-fine-tuned |        73.68           |     18:0      |             |
| Llama-3.1-8B-Instruct-fine-tuned |        69.74           |     23:0      |             |
|       gpt-4o              |        69.74           |     24:0      |    0.12     |
|       gpt-4o-mini         |        68.42           |     20:0      |    0.07     |
| gemini-1.5-flash-latest   |        68.42           |     18:2      |    0.07     |
| Llama-3.1-405B-Instruct   |        65.78           |     40:12     |             |
| Llama-3-70B-instruct      |        65.78           |     35:2      |             |
| Llama-3-8B-instruct       |        65.78           |     31.34     |             |
| gemini-1.5-pro-latest     |        64.47           |     34:40     |             |
| gpt-4-1106-preview        |        64.47           |     27:56     |    3.04     |
|          gpt-4            |        63.16           |     26:31     |    6.84     |
| claude-3-5-sonnet-20240620|        59.21           |     23:59     |    0.70     |
|  moa-gpt-3.5-turbo-0125   |        53.95           |     49:26     |             |
| gpt-4-0125-preview        |        53.94           |     34:40     |             |
|   patched-coder-7b        |        51.31           |     45.20     |             |
|  patched-coder-34b        |        46.05           |     33:58     |    0.87     |
|    patched-mix-4x7b       |        46.05           |     60:00+    |    0.80     |
|      Mistral-Large        |        40.80           |     60:00+    |             |
|       Gemini-pro          |        39.47           |     16:09     |    0.23     |
|      Mistral-Medium       |        39.47           |     60:00+    |    0.80     |
|      Mixtral-Small        |        30.26           |     30:09     |             |
|   gpt-3.5-turbo-0125      |        28.95           |     21:50     |             |
|  claude-3-opus-20240229   |        25.00           |     60:00+    |             |
| Llama-3-8B-instruct.Q4_K_M|        21.05           |     60:00+    |             |
|      Gemma-7b-it          |        19.73           |     36:40     |             |
|   gpt-3.5-turbo-1106      |        17.11           |     13:00     |    0.23     |
| Codellama-70b-Instruct    |        10.53           |     30.32     |             |
| CodeLlama-34b-Instruct    |         7.89           |     23:16     |             |

The price is calcualted by assuming 1000 input and output tokens per call as all examples in the dataset are < 512 tokens (OpenAI cl100k_base tokenizer). 

Some models timed out during the run or had intermittent API errors. We try each example 3 times in such cases. This is why some runs are reported to be longer than 1 hr (60:00+ mins).

If you want to add your model to the leaderboard, you can send in a PR to this repo with the log file from the evaluation run.