Disclaimer: This is a research prototype. The code and data in this repository are provided as-is, without any warranty or guarantee of correctness. Use at your own risk.
We tasks 7 LLMs with producing specifications for 20 data formats. results/ contains the generated files---without any annotations on whether these files compile.
-
Install necessary dependencies (pip packages for the LLM libraries)
-
Install the runtime libraries for the DSLs
- Install Kaitai Struct's compiler
- Download Apache Daffodil binaries
- Build the Spicy tools from Source: Spicy Documentation
- Build Galois' DaeDaLus tooling from source: DaeDaLus
- Clone the repository and build Hammer
git clone https://github.com/UpstandingHackers/hammer.git cd hammer scons sudo scons install -
Add API keys as environment variables.
- Together API to invoke the Llama and Deepseek LLMs.
- OpenAI
- Claude
- Gemini
export GOOGLE_API_KEY=""
export ANTHROPIC_API_KEY=""
export OPENAI_API_KEY=""
export TOGETHER_API_KEY=""
Together AI and Gemini are compatible with OpenAI's Python library.
-
Dockerfilebuilds a docker image with all of these compilers or executables installed. -
options.jsoncontains the paths of the various DDL executables and the list of formats and their specification versions. -
test.dbcontains the final database with two entire runs, where the run labeled "888" was the most recent run. -
Run the script to generate the DSLs using LLM queries
python3 create_dsls.py -
To generate the set of figures and tables used in the paper
python3 analyzer.py -
Test the generated library code by running a corpus of files through them. This command needs a folder containing files per format.
python3 compare-parsers.py
This work was supported in part by DOE NETL (DE-CR0000017) and the ARPA-H DIGIHEALS (Contract No. SP4701-23-C-0089). The views, opinions, and/or findings expressed are those of the author(s) and should not be interpreted as representing the official views or policies of DOE, ARPA-H, or the U.S. Government.