Release notes



  • Support for Python 3.12.

  • A new Sim4Ansatz based on the Sim et al. paper [SJA2019].

  • A new argument in for specifying an early_stopping_criterion other than validation loss.

  • A new argument collapse_noun_phrases in methods of CCGParser and CCGTree classes (for example, see CCGParser.sentence2diagram()) that allows the user to maintain noun phrases in the derivation or collapse them into nouns as desired.

  • Raised meaningful exception when users try to convert to/from DisCoPy 1.1.0


  • An internal refactoring of module backend.drawing in view of planned new features.

  • Updated random number generation in TketModel by using the recommended numpy.random.default_rnd() method.


  • Handling of possible empty Bra s and Ket s during conversion from DisCoPy.

  • Fixed a bug in JIT compilation of mixed circuit evaluations.



  • A new integrated backend that replaces DisCoPy, which until now was providing the low-level functionality of lambeq. The new backend offers better performance, increased stability, faster training speeds, and a simplified high-level interface to the user. The new backend consists of the following sub-modules:

  • BobcatParser: Added a special case for adjectival conjunction in tree translation.

  • TreeReader: Diagrams now are created straight from the CCGTree.

  • CCGRule apply method: Added apply() method to class CCGRule.


  • Diagram-level rewriters: Rewrite functions remove_cups() and remove_swaps() are now refactored as diagram-level rewriters, RemoveCupsRewriter and RemoveSwapsRewriter correspondingly.

  • Extra whitespace is now ignored in the Tokeniser.



  • Removed CCGTree.to_biclosed_diagram() and references to discopy.biclosed. Now CCG trees are directly converted into string diagrams, without the extra step of storing the derivation in a biclosed form.

  • CCGRule: Removed replace_cat_result() and added resolve().


This update features contributions from participants in unitaryHACK 2023:

  • Two new optimisers:
  • A new rewrite rule for handling unknown words. (credit: WingCode)

Many thanks to all who participated.

This update also contains the following changes:


  • DiagramRewriter is a new class that rewrites diagrams by looking at the diagram as a whole rather than by using rewrite rules on individual boxes. This includes an example UnifyCodomainRewriter which adds an extra box to the end of diagrams to change the output to a specified type. (credit: A.C.E07)

  • Added an early stopping mechanism to Trainer using the parameter early_stopping_interval.


  • In PennyLaneModel, SymPy symbols are now substituted during the forward pass so that gradients are back-propagated to the original parameters.

  • A pickling error that prevented CCG trees produced by BobcatParser from being unpickled has been fixed.



  • Support for DisCoPy >= 1.1.4 (credit: toumix).
    • replaced discopy.rigid with discopy.grammar.pregroup everywhere.

    • replaced discopy.biclosed with discopy.grammar.categorial everywhere.

    • Use Diagram.decode to account for the change in contructor signature Diagram(inside, dom, cod).

    • updated attribute names that were previously hidden, e.g. ._data becomes .data.

    • replaced diagrammatic conjugate with transpose.

    • swapped left and right currying.

    • dropped support for legacy DisCoPy.

  • Added CCGType class for utilisation in the biclosed_type attribute of CCGTree, allowing conversion to and from a discopy categorial object using discopy() and from_discopy() methods.

  • CCGTree: added reference to the original tree from parsing by introducing a metadata field.


  • Internalised DisCoPy quantum ansätze in lambeq.

  • IQPAnsatz now ends with a layer of Hadamard gates in the multi-qubit case and the post-selection basis is set to be the computational basis (Pauli Z).


  • Fixed a bottleneck during the initialisation of the PennyLaneModel caused by the inefficient substitution of Sympy symbols in the circuits.

  • Escape special characters in box labels for symbol creation.

  • Documentation: fixed broken links to DisCoPy documentation.

  • Documentation: enabled sphinxcontrib.jquery extension for Read the Docs theme.

  • Fixed disentangling RealAnsatz in extend-lambeq tutorial notebook.

  • Fixed model loading in PennyLane notebooks.

  • Fixed typo in SPSAOptimizer (credit: Gopal-Dahale)


  • Removed support for Python 3.8.



  • Added example and tutorial notebooks to tests.

  • Dependencies: pinned the maximum version of Jax and Jaxlib to 0.4.6 to avoid a JIT-compilation error when using the NumpyModel.


  • Documentation: fixed broken DisCoPy links.

  • Fixed PyTorch datatype errors in example and tutorial notebooks.

  • Updated custom ansätze in tutorial notebook to match new structure of CircuitAnsatz and TensorAnsatz.



  • Support for hybrid quantum-classical models using the PennyLaneModel. PennyLane is a powerful QML library that allows the development of hybrid ML models by hooking numerically determined gradients of parametrised quantum circuits (PQCs) to the autograd modules of ML libraries like PyTorch or TensorFlow.

  • Add lambeq-native loss functions LossFunction to be used in conjunction with the QuantumTrainer. Currently, we support the CrossEntropyLoss, BinaryCrossEntropyLoss, and the MSELoss loss functions.

  • Python 3.11 support.

  • An extensive NLP-101 tutorial, covering basic definitions, text preprocessing, tokenisation, handling of unknown words, machine learning best practices, text classification, and other concepts.


  • Improve tensor initialisation in the PytorchModel. This enables the training of larger models as all parameters are initialised such that the expected L2 norm of all output vectors is approximately 1. We use a symmetric uniform distribution where the range depends on the output dimension (flow) of each box.

  • Improve the fail-safety of the BobcatParser model download method by adding hash checks and atomic transactions.

  • Use type union expression | instead of Union in type hints.

  • Use raise from syntax for better exception handling.

  • Update the requirements for the documentation.


  • Fixed bug in SPSAOptimizer triggered by the usage of masked arrays.

  • Fixed test for NumpyModel that was failing due to a change in the behaviour of Jax.

  • Fixed brittle quote-wrapped strings in error messages.

  • Fixed 400 response code during Bobcat model download.

  • Fixed bug where CircuitAnsatz would add empty discards and postselections to the circuit.


  • Removed install script due to deprecation.



  • Improved the performance of NumpyModel when using Jax JIT-compilation.

  • Dependencies: pinned the required version of DisCoPy to 0.5.X.


  • Fixed incorrectly scaled validation loss in progress bar during model training.

  • Fixed symbol type mismatch in the quantum models when a circuit was previously converted to tket.




  • Dependencies: bumped the minimum versions of discopy and torch.

  • IQPAnsatz now post-selects in the Hadamard basis.

  • PytorchModel now initialises using xavier_uniform.

  • CCGTree.to_json() can now be applied to None, returning None.

  • Several slow imports have been deferred, making lambeq much faster to import for the first time.

  • In CCGRule.infer_rule(), direction checks have been made explicit.

  • UnarySwap is now specified to be a unaryBoxConstructor.

  • BobcatParser has been refactored for easier use with external evaluation tools.

  • Documentation: headings have been organised in the tutorials into subsections.




  • Added a strict pregroups mode to the CLI. With this mode enabled, all swaps are removed from the output string diagrams by changing the ordering of the atomic types, converting them into a valid pregroup form as given in [Lam1999].

  • Adjusted the behaviour of output normalisation in quantum models. Now, NumpyModel always returns probabilities instead of amplitudes.

  • Removed the prediction from the output of the SPSAOptimizer, which now returns just the loss.


  • Added a “swapping” unary rule box to handle unary rules that change the direction of composition, improving the coverage of the BobcatParser.

  • Added a --version flag to the CLI.

  • Added a make_checkpoint() method to all training models.

  • Changed the WebParser so that the online service to use is specified by name rather than by URL.

  • Changed the BobcatParser to only allow one tree per category in a cell, doubling parsing speed without affecting the structure of the parse trees (in most cases).

  • Fixed the parameter names in CCGRule, where dom and cod had inadvertently been swapped.

  • Made the linting of the codebase stricter, enforced by the GitHub action. The flake8 configuration can be viewed in the setup.cfg file.


  • Fix a bug that caused the BobcatParser and the WebParser to trigger an SSL certificate error using Windows.

  • Fix false positives in assigning conjunction rule using the CCGBankParser. The rule , + X[conj] -> X[conj] is a case of removing left punctuation, but was being assigned conjunction erroneously.

  • Add support for using jax as backend of tensornetwork when setting use_jit=True in the NumpyModel. The interface is not affected by this change, but performance of the model is significantly improved.


  • Fix a bug that raised a dtype error when using the TketModel on Windows.

  • Fix a bug that caused the normalisation of scalar outputs of circuits without open wires using a QuantumModel.

  • Change the behaviour of spiders_reader such that the spiders decompose logarithmically. This change also affects other rewrite rules that use spiders, such as coordination and relative pronouns.


  • CCGRule: Add symbol() method that returns the ASCII symbol of a given CCG rule.

  • CCGTree: Extend deriv() method with CCG output. It is now capable of returning standard CCG diagrams.

  • Command-line interface: Add CCG mode. When enabled, the output will be a string representation of the CCG diagram corresponding to the CCGTree object produced by the parser, instead of a DisCoPy diagram or circuit.

  • Documentation: Add a troubleshooting page.


  • Add support for Python 3.10.

  • Unify class hierarchies for parsers and readers: CCGParser is now a subclass of Reader and placed in the common package text2diagram. The old packages reader and ccg2discocat are no longer available. Compatibility problems with previous versions should be minimal, since from Release 0.2.0 and onwards all lambeq classes can be imported from the global namespace.

  • Add CurryRewriteRule, which uses map-state duality in order to remove adjoint types from the boxes of a diagram. When used in conjunction with normal_form(), this removes cups from the diagram, eliminating post-selection.

  • The Bobcat parser now updates automatically when new versions are made available online.

  • Update grammar file of Bobcat parser to avoid problems with conflicting unary rules.

  • Allow customising available root categories for the parser when using the command-line interface.


  • A new Checkpoint class that implements pickling and file operations from the Trainer and Model.

  • Improvements to the training module, allowing multiple diagrams to be accepted as input to the SPSAOptimizer.

  • Updated documentation, including sub-package structures and class diagrams.


  • A new state-of-the-art CCG parser based on [SC2021], fully integrated with lambeq, which replaces depccg as the default parser of the toolkit. The new Bobcat parser has better performance, simplifies installation, and provides compatibility with Windows (which was not supported due to a depccg conflict). depccg is still supported as an alternative external dependency.

  • A training package, providing a selection of trainers, models, and optimizers that greatly simplify supervised training for most of lambeq’s use cases, classical and quantum. The new package adds several new features to lambeq, such as the ability to save to and restore models from checkpoints.

  • Furthermore, the training package uses DisCoPy’s tensor network capability to contract tensor diagrams efficiently. In particular, DisCoPy 0.4.1’s new unitary and density matrix simulators result in substantially faster training speeds compared to the previous version.

  • A command-line interface, which provides most of lambeq’s functionality from the command line. For example, lambeq can now be used as a standard command-line pregroup parser.

  • A web parser class that can send parsing queries to an online API, so that local installation of a parser is not strictly necessary anymore. The web parser is particularly helpful for testing purposes, interactive usage or when a local parser is unavailable, but should not be used for serious experiments.

  • A new pregroups package that provides methods for easy creation of pregroup diagrams, removal of cups, and printing of diagrams in text form (i.e. in a terminal).

  • A new TreeReader class that exploits the biclosed structure of CCG grammatical derivations.

  • Three new rewrite rules for relative pronouns [SCC2014a] [SCC2014b] and coordination [Kar2016].

  • Tokenisation features have been added in all parsers and readers.

  • Additional generator methods and minor improvements for the CCGBankParser class.

  • Improved and more detailed package structure.

  • Most classes and functions can now be imported from lambeq directly, instead of having to import from the sub-packages.

  • The circuit and tensor modules have been combined into an lambeq.ansatz package. (However, as mentioned above, the classes and functions they define can now be imported directly from lambeq and should continue to do so in future releases.)

  • Improved documentation and additional tutorials.


  • Add URLs to the setup file.

  • Fix logo link in README.

  • Fix missing version when building docs in GitHub action.

  • Fix typo in the description keyword of the setup file.


  • Update install script to use PyPI package.

  • Add badges and documentation link to the README file.

  • Add lambeq logo and documentation link to the GitHub repository.

  • Allow documentation to get the package version automatically.

  • Add keywords and classifiers to the setup file.

  • Fix: Add lambeq.circuit module to top-level lambeq package.

  • Fix references to license file.


The initial release of lambeq, containing a lot of core material. Main features:

  • Converting sentences to string diagrams.

  • CCG parsing, including reading from CCGBank.

  • Support for the depccg parser.

  • DisCoCat, bag-of-words, and word-sequence compositional models.

  • Support for adding new compositional schemes.

  • Rewriting of diagrams.

  • Ansätze for circuits and tensors, including various forms of matrix product states.

  • Support for JAX and PyTorch integration.

  • Example notebooks and documentation.