Skip to content


Switch branches/tags


Failed to load latest commit information.


Roosterize is a tool for suggesting lemma names in verification projects that use the Coq proof assistant. The tool leverages neural networks that take serialized Coq lemma statements and elaborated terms as input; see the Technique section below.



You can install Roosterize from source code by cloning this GitHub repository and setting up the dependencies following steps 1 & 2. (Alternatively, you can download the a binary distribution which already contains the Python dependencies, and then you only need step 1.)

git clone
cd roosterize

1. Installation of OCaml, Coq, and SerAPI

We strongly recommend installing the required versions of OCaml, Coq, and SerAPI via the OPAM package manager, version 2.0.7 or later.

To set up the OPAM-based OCaml environment, use:

opam switch create roosterize 4.07.1
opam switch roosterize
eval $(opam env)

Then, install Coq and SerAPI, pinning them to avoid unintended upgrades:

opam update
opam pin add coq 8.10.2
opam pin add coq-serapi 8.10.0+0.7.1

2. Installation of PyTorch and Python libraries

We strongly recommend installing the required versions of Python and PyTorch using Conda.

To set up the Conda environment, use one of the following command suitable for your operating system and whether you want to use it on a CPU or GPU.

  • Linux, CPU:
conda env create --name roosterize --file conda-envs/cpu.yml
  • Linux, GPU w/ CUDA 10.0:
conda env create --name roosterize --file conda-envs/gpu-cuda10.yml
  • Linux, GPU w/ CUDA 9.0:
conda env create --name roosterize --file conda-envs/gpu-cuda9.yml
  • Mac, CPU:
conda env create --name roosterize --file conda-envs/mac-cpu.yml

Finally, activate the Conda environment before using Roosterize:

conda activate roosterize

Installation of trained models

Next, you need to obtain a pre-trained model that capture naming conventions. The default pre-trained model, which was trained using our corpus and follows the conventions used in the Mathematical Components family of projects, can be obtained by running the command:

./bin/roosterize download_global_model

The model will be downloaded to $HOME/.roosterize/. To use a different model (that we released or you trained), simply put it in $HOME/.roosterize/.


To use Roosterize on a Coq verification project, you first need to build the Coq project using a command provided by the project (usually make). Then, run this command to get the lemma name suggestions for the lemmas in a Coq document (.v file):

python -m roosterize.main suggest_naming --file=PATH_TO_FILE

Roosterize automatically infers the root directory of the project by finding _CoqProject, and reads _CoqProject to infer the SerAPI command line options (for mapping logical paths to directories, see SerAPI's documentation). If you don't have a _CoqProject file, you need to provide an additional argument --project_root=PATH_TO_PROJECT_ROOT.

The above command extracts all lemmas from the file, uses the pre-trained model to generate likely lemma names for each lemma, and finally prints the lemma name update suggestions, i.e., the generated lemma names that are different from to the existing ones. See an example suggestion report.

For other usages and command line interfaces of Roosterize, please check the help:

python -m roosterize.main help


Roosterize learns and suggests lemma names using neural networks that take serialized Coq lemma statements and elaborated terms as input. For example, the Coq lemma sentence

Lemma mg_eq_proof L1 L2 (N1 : mgClassifier L1) : L1 =i L2 -> nerode L2 N1.

is serialized into the following tokens (simplified):

(Sentence((IDENT Lemma)(IDENT mg_eq_proof)(IDENT L1)(IDENT L2)
  (KEYWORD"(")(IDENT N1)(KEYWORD :)(IDENT mgClassifier)

and the corresponding elaborated term (simplified):

(Prod (Name (Id char)) ... (Prod (Name (Id L1)) ...
 (Prod (Name (Id L2)) ... (Prod (Name (Id N1)) ...
  (Prod Anonymous (App (Ref (DirPath ((Id ssrbool) (Id ssr) (Id Coq))) (Id eq_mem)) ...
   (Var (Id L1)) ... (Var (Id L2)))
  (App (Ref (DirPath ((Id myhill_nerode) (Id RegLang))) (Id nerode)) ...
   (Var (Id L2)) ... (Var (Id N1))))))))

The diagram below illustrates Roosterize's neural network architecture, as applied to this example:

Our research paper outlines the design of Roosterize, and describes an evaluation on a corpus of serialized Coq code derived from the Mathematical Components family of projects. The training, validation, and testing sets of Coq files from the corpus used in the evaluation are defined in the training directory.

If you have used Roosterize in a research project, please cite the research paper in any related publication:

  author = {Nie, Pengyu and Palmskog, Karl and Li, Junyi Jessy and Gligoric, Milos},
  title = {Deep Generation of {Coq} Lemma Names Using Elaborated Terms},
  booktitle = {International Joint Conference on Automated Reasoning},
  pages = {97--118},
  doi = {10.1007/978-3-030-51054-1_6},
  year = {2020},