DialogStudio: a unified dialog collection that keeps the originals

Unified without flattening, and that is the design decision

DialogStudio is a large collection and unified dialog datasets, and the sentence that explains how it works is precise. DialogStudio unified each dataset while preserving its original information, and this aids in supporting research on both individual datasets and Large Language Model training.

So the unification is a format change, not a merge of the underlying conversations. That distinction is the whole point of the project, because a collection that flattens everything into one schema is convenient for training and useless for anyone who wants to work on the original dataset.

The project description says the same thing more fully, towards the richest and most diverse unified dataset collection and instruction-aware models for conversational AI, and the paper is linked from the readme alongside the Hugging Face dataset, the model section and a Twitter link.

The data are downloadable through Hugging Face, and examples for each dataset are provided in the repository itself. For granular and category-specific details, the readme points at the individual folders corresponding to each category, such as the MULTIWOZ2_2 dataset under the task-oriented-dialogues category.

The full list of available datasets is a separate file, Dataset_Stats.csv, and there is also a spreadsheet linked in a comment that is described as the check for all supported datasets.

The categories are six, and they cover the range of what a dialogue dataset is used for. Knowledge-Grounded-Dialogues, Natural-Language-Understanding, Open-Domain-Dialogues, Task-Oriented-Dialogues, Dialogue-Summarization and Conversational-Recommendation-Dialogs. The top level of the repository has a directory for each, plus a code/ directory, a figures/ directory and a stats/ directory.

The project is Apache-2.0 licensed, with the licence in a LICENSE.txt file, and the last push was on 2026-06-02 with a single release tag, version 1.0.1 from August 2023.

Loading a dataset takes two lines and needs no login

The loading section is short, and one of its changes is worth knowing about because it removed a barrier.

You load any dataset in DialogStudio from the Hugging Face hub by claiming the dataset name, which is exactly the dataset folder name. So the folder name in the repository is the identifier you pass to the loader, and there is no separate mapping table to consult.

The example loads MULTIWOZ2_2 from the task-oriented-dialogues category:

from datasets import load_dataset

dataset = load_dataset('Salesforce/dialogstudio', 'MULTIWOZ2_2')

And the output structure shows what you get back, which is a DatasetDict with train, validation and test splits:

DatasetDict({
    train: Dataset({
        features: ['original dialog id', 'new dialog id', 'dialog index', 'original dialog info', 'log', 'prompt', 'external knowledge non-flat', 'external knowledge', 'dst knowledge', 'intent knowledge'],
        num_rows: 8437
    })
    validation: Dataset({

The feature list is the interesting part, because it shows what preserving the original information means in practice. Alongside the original dialog id there is a new dialog id, and there is a dialog index and the original dialog info. Then come the log, the prompt, external knowledge in both a non-flat and a flat form, dst knowledge, and intent knowledge.

So the unified schema keeps a pointer back to the source conversation, adds a stable identifier of its own, and carries the dialogue history and prompt that an instruction-tuned model would consume, alongside the external knowledge the task-oriented datasets bring with them.

The external knowledge appears twice, once non-flat and once flat. The non-flat form preserves the original structure, and the flat form is the version a model can actually use, which is the second half of the preserve-and-unify promise: nothing is thrown away, and the thing a trainer needs is also there.

The login requirement was removed in the version 1.0.1 release, which is listed among the minor issues that release resolved along with the resolution of minor issues in a few dialogues, the addition of prompts for selected knowledge-grounded datasets, and updates to the SODA and ShareGPT datasets.

There is a comment in the readme about agreeing to licences on the Hugging Face hub and about logging in locally, with the login command shown. It is commented out rather than presented as a step, which matches the release note that the login requirement was removed.

Quality scoring with a script you can read

The collection comes with its own quality assessment, and the method is stated plainly enough to be useful.

DialogStudio evaluates dialogue quality based on six critical criteria, namely Understanding, Relevance, Correctness, Coherence, Completeness, and Overall Quality. Each criterion is scored on a scale of 1 to 5, with the highest scores reserved for exceptional dialogues.

Given the number of datasets incorporated, the assessment was done with gpt-3.5-turbo across 33 distinct datasets, and the script used for the evaluation is linked from the readme, sitting in the code/ directory of the repository.

So the scores are model-generated rather than human, applied at the dataset level rather than at the dialogue level, and the script that produced them is published. That last point is what separates this from a quality score you have to take on trust.

The results are presented as a figure, and the readme adds that they intend to release evaluation scores for individually selected dialogues in the upcoming period. So the committed scores are per dataset, and per-dialogue scores were still to come as of the release.

There is a general statistics figure as well, a summary of the statistics associated with DialogStudio, and a stats/ directory at the repository root for the underlying numbers.

The AI_ETHICS.md file at the repository root is the piece that makes the gpt-3.5-turbo approach defensible rather than merely expedient. Scoring a large collection of dialogues with a commercial model is a decision with implications for what the scores mean, and documenting that decision is what allows a user to decide whether to weight the scores at all.

The readme also links a general statistics figure and the score figure, so the summary of the collection is a picture rather than a table, with the numeric detail in Dataset_Stats.csv and the stats/ directory.

Instruction-tuned models trained on a subset of the collection

The model section is smaller than the dataset section, and the news entries describe what was released.

In August 2023 the version 1.0 models were uploaded, in three sizes: dialogstudio-t5-base-v1.0, dialogstudio-t5-large-v1.0 and dialogstudio-t5-3b-v1.0. The description says they were trained on a few selected DialogStudio datasets and more than 1000 general tasks.

That description is worth reading carefully, because it bounds what the models are. The T5 family is an encoder-decoder rather than a decoder-only architecture, so these are sequence-to-sequence models rather than chat models in the usual sense, and the general tasks are what give them instruction-following behaviour outside the dialogue domains the collection covers.

The phrase a few selected datasets is doing real work. The collection spans six categories and dozens of datasets, and the models were trained on a small selection of them alongside a large general task set. That is a deliberate choice, because training on everything would produce a model that is average across domains rather than good at the ones that matter, and the general tasks are there to keep the model useful as an instruction follower.

The release naming is consistent, with the dataset and the version in the name, so a model name tells you both what it was trained on and which release it belongs to.

The model link in the readme header points to a section of the same document rather than to a separate page, and the newest news entry points elsewhere entirely. In March 2024 the announcement is that xLAM has been updated for AI Agent work, with a link to the xLAM repository for the latest data and models relevant to that. The DialogStudio collection itself has not moved on since the version 1.0.1 release in August 2023.

There is also a March 2024 entry about Hugging Face dataset viewer issues, which is a practical note rather than a research one. The fix is to refer to the repository for a view of each dataset, where five converted examples are provided along with five original examples under each data folder. The ShareGPT folder is given as the example, containing a converted_examples.json file and an original_example.json file.

That arrangement, converted and original side by side, is the same preserve-and-unify idea expressed as files you can diff.

What the repository is for, and where its limits are

The release history is short, which tells you something about how this project ended.

There is one tag, version 1.0.1 from August 2023. The initial release was in July 2023, described as the largest unified Dialog dataset collection. The version 1.0 models followed in August 2023, and the version 1.0.1 release fixed minor issues, added prompts and updated two datasets. Then the last two news entries are from March 2024, and they point elsewhere: one to xLAM for AI Agent work, and one to a workaround for the dataset viewer.

So the collection itself stopped moving in 2023, and the repository has since become a reference point that other work is referenced from, rather than a growing collection. The last push was on 2026-06-02, which is a maintenance push rather than a release.

For anyone planning to use it, that is a good position to be in. A dataset collection that has not changed in two years is stable, and the per-dataset examples in the repository let you check the shape of the data before you download anything.

The practical limits are the ones the readme states. The quality scores are generated by a model rather than by people, at the dataset level, with per-dialogue scores still pending, so they should be treated as a coarse signal. The datasets are claimed on the Hugging Face hub, so each one has a licence agreement to accept, and the version 1.0.1 release removed the need for a login but did not remove the licence step. And the unified schema is a conversion of the original data, so a project that needs the raw format still needs the original dataset.

The repository carries the usual policy files for a project of this size, an AI_ETHICS.md, a CODE_OF_CONDUCT.md, a CONTRIBUTING.md, a SECURITY.md and a CODEOWNERS file, and the paper is committed alongside the readme as a PDF in the repository root, so the methodology travels with the data rather than living only on a preprint server.

Editorial conclusion

DialogStudio suits a researcher who needs a broad dialog corpus in one schema and who also cares about what the original datasets looked like, since the converted and original forms sit side by side and the dataset examples are in the repository. It does not suit you if you need per-dialogue quality judgements, since those were still pending at the last release, and the scores are model-generated at the dataset level, so they are a coarse filter rather than a ranking. Before you build on it, read the AI_ETHICS.md file, because that is where the consequences of scoring with a commercial model are set out, and check that the licence agreement on the hub is acceptable to you, since every dataset is claimed separately.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. README
4. Releases
5. salesforce/DialogStudio on GitHub