AlibabaResearch/DAMO-ConvAI: a monorepo of research code, not a conversational AI framework

What DAMO-ConvAI actually is: a container for separate research projects

The repository README describes DAMO-ConvAI as "the official repository which contains the codebase for Alibaba DAMO Conversational AI". That sentence is the whole scope statement. The top level of the repository is a list of directories, and each one is a distinct project: r2sql, s2sql, sunsql, bird, space-1, space-2, space-3, doc2bot, api-bank, dater, dial2vec, dialogue-cse, spokenwoz, oltqa, cgodial, graphix, metaretriever, mmevol, deep-thinking, and others. There is no shared package, no shared entry point, and no top-level setup script.

The README's News section is a publication log rather than a changelog. It records paper acceptances and leaderboard results: S2SQL accepted at ACL 2022 and first rank on the Spider leaderboard, R2SQL at AAAI 2021 with first rank on SparC and CoSQL, SPACE 1.0 at AAAI 2022, SPACE 3.0 at SIGIR 2022, BIRD-SQL accepted at NeurIPS 2023 as a spotlight, and a run of ACL, EMNLP, KDD, LREC-COLING and SigDial acceptances through 2024. That is the useful signal here. This is a research artifact archive, and the unit of adoption is the individual directory, not the repository.

Who it is for follows from that. If you are reproducing a paper, comparing against a published baseline, or looking for how a specific model was implemented, the relevant subdirectory is the thing you want. If you are looking for a dialog framework to put behind a product, this repository does not present itself that way.

The repository layout is the architecture

There is no architecture document because there is no single system. The data flow is per project and defined inside each directory. What the repository layout tells you is that projects are grouped by research area rather than by dependency: the space-1, space-2 and space-3 directories are successive versions of the SPACE line, r2sql, s2sql, sunsql and bird sit near each other as text-to-SQL work, and spokenwoz, dater, cgodial and doc2bot are dialog datasets and models.

That grouping is a reading aid, not a build graph. Nothing in the README states that any directory depends on another, and there is no shared configuration, no shared model registry and no common CLI. If you want to run two of these projects, treat them as two independent checkouts that happen to share a parent directory.

The practical consequence is that you cannot reason about DAMO-ConvAI as a whole. Questions like "what Python version does it need" or "what is the inference API" have no repository-level answer. Each subdirectory carries its own answer, or does not carry one at all. That is a normal shape for a lab repository and an unusual shape for anything you would vendor into a service.

Installing it: there is no repository-level install

The README does not give installation steps, and it does not give a package name. There is no pip install line for DAMO-ConvAI as a whole, no setup.py or pyproject.toml at the top level in the repository listing, and no release artifacts. The README is a title, a license block and a news list. So the honest first step is to clone the repository and go into the subdirectory for the project you actually want.

git clone https://github.com/AlibabaResearch/DAMO-ConvAI.git
cd DAMO-ConvAI
ls

The ls output is the directory list quoted above: r2sql, s2sql, space-1, bird, doc2bot, api-bank and the rest. From here, the README of the subdirectory you picked is your installation guide, not this one.

cd s2sql
ls

What you see in that second listing depends entirely on which project you chose, and the repository README does not describe the contents of any of them. If the subdirectory has no README and no requirements file, you are reading source code to work out how to run it. Plan for that possibility before you commit time to a project.

The MIT licence text is reproduced in full in the README and applies to the repository. That is the one piece of repository-wide information you can rely on without opening a subdirectory.

Subdirectories are the real unit of work, and they are uneven

The main limitation is not a bug, it is the shape of the thing. A monorepo of research projects accumulates directories at different levels of polish, and nothing at the top level tells you which is which. The README's news entries are evidence that a paper was published, not that the code in the matching directory runs end to end today.

Concretely, if you pick a directory and find no dependency list, no documented entry point and no example command, you have three options: read the source, find the paper and reconstruct the setup, or pick a different directory. None of those is unusual for research code, but all of them cost time that a README would have saved.

A second constraint is version drift. Projects accepted at AAAI 2021 and projects accepted at LREC-COLING 2024 sit in the same repository, and the frameworks they were written against are years apart. There is no stated policy in the README about updating older directories to newer library versions. Assume each directory is frozen at the environment its authors used.

This is also why DAMO-ConvAI is the wrong tool for a production dialog system. There is no supported inference path, no versioning of the kind a dependency needs, and no compatibility promise across directories. Using it in production means adopting one subdirectory as your own code and maintaining it yourself.

DAMO-ConvAI compared with Hugging Face Transformers

The obvious alternative for most people reading this is Hugging Face Transformers, and the difference is not model quality. It is what the two things are. Transformers is a library: one package, a documented API, versioned releases, and a model hub you pull from at runtime. DAMO-ConvAI is a set of repositories-in-a-repository, with no package and no API surface that spans projects.

If your goal is to run a pretrained dialog or text-to-SQL model, Transformers gives you an install command and a loading call. If your goal is to understand or reproduce a specific method from an Alibaba DAMO paper, Transformers does not contain that method, and DAMO-ConvAI does. That is the split. One is a runtime dependency, the other is a reference.

A narrower alternative for text-to-SQL work specifically is to go to the leaderboards the README itself cites, Spider, SparC and CoSQL, and take the implementation from whichever entry matches your constraints. DAMO-ConvAI's r2sql and s2sql are among the entries those leaderboards list, so this is not either-or: the repository is one source among several on the same benchmark.

Maintenance, licence and what upgrading means here

The repository is not archived, and the last push was on 2026-06-10. That is a fact about the repository as a whole and says nothing about any individual directory, which may have been touched years earlier or not at all. There are no retrieved releases, so there is no version number to pin and no changelog to read between versions.

Upgrading, in this repository, means pulling the default branch. Because there is no packaging, there is no dependency resolver to tell you what changed; you diff the subdirectory you care about and read the commit history for that path. If you have forked or vendored a subdirectory, that diff is your upgrade process, and you own the merge.

The licence is MIT, stated in the README and present as a LICENSE file at the top level, with copyright attributed to Alibaba Research and a 2022 date in the notice. MIT is permissive: it allows use, modification and redistribution provided the copyright notice and permission notice are included. Two things the README does not address, and which you should check for your own situation rather than assume: whether model weights or datasets referenced by a subdirectory carry separate terms, and whether third-party code inside a subdirectory carries its own licence. This is not legal advice; read the LICENSE file and any per-directory notices before you redistribute anything.

Editorial conclusion

Adopt DAMO-ConvAI if you are reproducing a specific paper or need a reference implementation for text-to-SQL or dialog modelling, and check that the subdirectory you need contains its own README and requirements before you plan any work around it. Do not adopt it as a single dependency for a product: there is no package, no unified API, and no install command in the repository README. Verify first that the subdirectory for your target project exists, that it carries its own setup instructions, and which Python version and dependencies it expects, since nothing at the top level states this.

Frequently asked questions

Official sources

1. AlibabaResearch/DAMO-ConvAI on GitHub
2. Issues
3. License: MIT
4. README