Align ML.NET labels with dotnet/runtime

[rosebyte]

This proposal introduces runtime-style area-* labels in dotnet/machinelearning, aligns shared label names, and removes duplicate or ambiguous labels.

Following feedback, this version starts with broad areas. The previous version proposed 38 labels, including subdivisions of Core, DataFrame, Vision, AutoML, and Infrastructure. This version proposes 19 top-level areas. Subareas can be added later if they have separate owners or support a recurring query that cannot be served by other labels.

It does not copy runtime's full label set. Most runtime labels are specific to CLR, JIT, servicing, operating systems, or architectures.

Current state

Snapshot from 4 September 2026:

Measure	ML.NET	Runtime
Total labels	70	316
area-* labels	1	146

area-Infrastructure has already been created and is applied to five open issues.

Runtime's issue guide recommends one area label per issue, and its pull-request guide describes assigning a single area to each pull request.

This is a strong default, not an absolute rule. A live check on 2 September 2026 found multiple areas on 4 of 565 open runtime issues and 3 of 1,000 sampled open pull requests.

For ML.NET, an item should normally have one primary area. Add another area when ownership is genuinely shared, not merely because several technologies are mentioned.

The supplied ml_net-area_labels.md currently contains 1,011 issue assignments. The repository currently has 808 open issues:

• 796 open issues occur in the file.
• 215 issues in the file have since closed.
• 12 newer open issues are absent.

The mapping needs refreshing before it is applied.

Proposed area labels

Area label	Scope
area-Core	MLContext, public API, prediction, evaluation, data loading, model persistence, IDataView, schema, and other core platform work
area-Native	CpuMath SIMD, MKL, oneDAL, SymSGD, native libraries, and hardware or RID plumbing
area-MAML	Legacy MAML CLI, entry-point graph, sweeper, and result processor
area-DataFrame	Microsoft.Data.Analysis, including columns, operations, and input/output
area-Transforms	Normalisation, categorical and text featurisation, LDA, embeddings, mappings, and feature selection
area-Trainers	Standard linear and GLM trainers, KMeans, PCA, Naive Bayes, OVA, and ensemble meta-trainers
area-Trees	FastTree, FastForest, GAM, LightGBM, and their native engine
area-Recommender	Matrix factorisation, field-aware factorisation machines, and libmf
area-TimeSeries	SSA forecasting, spike and change-point detection, SR-CNN, and time-series anomaly detection
area-Vision	Image loading and transforms, image classification, and object detection
area-TensorFlow	TensorFlow scoring, retraining, Keras, SavedModel, and TensorFlow interoperation
area-ONNX	OnnxTransformer scoring and ML.NET-to-ONNX export
area-Tokenizers	BPE, WordPiece, SentencePiece, Tiktoken, and tokenizer data packages
area-GenAI	Causal language-model pipeline and LLaMA, Phi, and Mistral support
area-TorchSharp	TorchSharp-based NLP and deep-learning components
area-AutoML	AutoML experiments, search and tuning, Model Builder, mlnet CLI, and code generation
area-Integration	ASP.NET Core dependency injection, PredictionEnginePool, and Microsoft.Extensions.ML
area-Infrastructure	Build, CI, Helix, packaging, dependencies, shared test infrastructure, and benchmark infrastructure
area-Meta	Roadmap, repository direction, governance, issue management, and other repository-wide work

Each area should have maintainers who recognise it as theirs. Otherwise it is only a category, not a routing label.

Treatment of the detailed taxonomy

The finer groups in ml_net-area_labels.md remain useful for understanding the backlog, but are not proposed as GitHub labels initially.

Detailed group	Initial GitHub label
area-core-*	area-Core
area-dataframe-*	area-DataFrame
area-vision-image	area-Vision
area-vision-classification	area-Vision
area-vision-objectdetection	area-Vision
area-vision-tensorflow	area-TensorFlow
area-automl-*	area-AutoML
area-infra-*	area-Infrastructure
area-docs	The affected component area, or area-Meta for repository-wide material, together with documentation

This avoids creating labels for distinctions such as Infrastructure Build, CI, Testing, and Performance when they currently have no separate ownership. Those distinctions can still be expressed through issue-type or qualifier labels.

Existing labels that can be renamed

Existing label	New label	Note
** NO MERGE **	NO-MERGE	Runtime name
API breaking change	breaking-change	Runtime name
P0	Priority:0	Runtime name
P1	Priority:1	Runtime name
P2	Priority:2	Runtime name
P3	Priority:3	Runtime name
perf	tenet-performance	Performance is a qualifier, not ownership
AutoML.NET	area-AutoML	Covers all proposed AutoML subdivisions
Microsoft.Data.Analysis	area-DataFrame	Covers all proposed DataFrame subdivisions
ModelBuilder	area-AutoML	Model Builder remains searchable by text and project
PredictionEngine	area-Core	Prediction is no longer a separate area
TensorFlow	area-TensorFlow	Direct component match
Time Series	area-TimeSeries	Direct component match
Tokenizers	area-Tokenizers	Direct component match
TorchSharp	area-TorchSharp	Direct component match

onnx can become area-ONNX after moving #7383 to area-Tokenizers.

Existing labels that should be merged

Existing labels	Keep
Awaiting User Input, need info	needs-author-action
up-for-grabs	help wanted
anomaly, Time Series	area-TimeSeries

Existing labels that need individual reclassification

These labels span several areas or describe a scenario rather than an owner.

Existing label	Treatment
API	Assign the component area and move API proposals to the appropriate api-* state
Build	Move infrastructure issues to area-Infrastructure; #5569 belongs to area-AutoML
Deep Learning	Split among TorchSharp, ONNX, Tokenizers, Vision, TensorFlow, GenAI, and AutoML
image	Split among Vision, TensorFlow, and AutoML
loadsave	Split among Core, DataFrame, and AutoML
command-line	Split between MAML and AutoML
test	Assign the component area and use a specific test issue type where useful
lightgbm	Usually Trees, but current outliers need review
NLP	Split among Transforms, TorchSharp, Tokenizers, ONNX, and AutoML
Hardware Support	Assign the component area and keep a platform qualifier if useful
Dynamic	Move current issues to Core
Evaluation	Move current issues to Core
Explainability	Split between Core and the relevant trainer area
oneDAL	Move current issues to Native
Spark	Assign the underlying component; keep a separate area only if Spark has an active owner
UWP	Assign the underlying component and keep a platform qualifier only if UWP remains supported
classification	Split among Trainers, AutoML, TorchSharp, and Vision
clustering	Move current issues to Trainers
ranking	Move current issues to Trees unless another owner is more appropriate
regression	Assign the relevant trainer or Vision area

Task labels such as task-classification or task-ranking can be added later if maintainers actually use that view of the backlog.

ML.NET-specific labels to keep

Label	Use
F#	F# language interoperation
usability	User-experience and discoverability work, if maintainers still query it
repo-health	Identifies the dashboard issue used by repository-health workflows

Labels to retire or remove

Label	Treatment
Azure AutoML	Retire; it has no open issues and one closed issue
code-sanitation	Replace with enhancement
nit	Use good first issue where appropriate
wontfix	Use GitHub's not planned close reason

A rename preserves historical associations. For labels that cannot be renamed directly, remove them only after their open issues have been reclassified.

Shared labels to keep

Label	Use or change
agentic-workflows	Agentic Workflow changes
api-approved	Approved API proposal
api-needs-work	API proposal requiring revision
api-ready-for-review	API proposal ready for review
blocking	Work that blocks something important
blocking-clean-ci	CI-blocking failure
bug	Product defect
community-contribution	Pull request submitted by a community member
documentation	Documentation-only bug or improvement
enhancement	Product improvement that does not add public API
Epic	Groups related work
good first issue	Small, well-scoped first contribution
help wanted	Actionable work open to external contributors
in-pr	An active pull request is expected to close the issue
Known Build Error	Known CI or Helix failure
needs-author-action	More information or action is required from the author
needs-further-triage	Maintainers need to make a deeper decision
no-recent-activity	Stale-item automation marker
question	Usage or product question; replace the current Further information is requested description
reduce-unsafe	Work reducing unsafe code
Security	Security-sensitive issue
untriaged	Not yet triaged by maintainers

Where names are shared with runtime, descriptions and colours can be aligned unless ML.NET has a reason to differ.

Labels worth adding now

Label	Use
needs-area-label	No area could be assigned confidently
api-suggestion	Early public API idea or discussion
feature-request	New product capability before a specific API proposal exists
blocked	Work cannot proceed because it depends on something else
test-bug	Defect in test code
test-enhancement	Improvement to test code or coverage

Other runtime labels, including cost, tenet, platform, servicing, and tracking labels, should be added only when ML.NET has a concrete use for them.

Repository references that need updating

File	Current references
.github/workflows/issue-triage.agent.md	perf, test, Build, need info; areas are mentioned only in comments
.github/workflows/issue-triage.agent.lock.yml	Generated allow-list containing the old names
.github/workflows/repo-health-check.md	P0 through P3, Awaiting User Input, and a statement that the repository has no area labels
.github/workflows/repo-health-check.lock.yml	Generated references to the old names
.github/health-baseline.md	Old priority names

If automated area labelling is wanted, runtime uses dotnet/issue-labeler with LABEL_PREFIX: "area-" and DEFAULT_LABEL: "needs-area-label". It should be trained only after the new area assignments are clean.

[github-actions]

🔍 No similar issues found with high confidence.

---

AI-powered similar issue detection

[matouskozak]

One thing I would consider is simplifying some sub-areas unless we really need the added granularity. E.g., I think area-Infrastructure is enough.

[rosebyte]

Good idea, I've reduced the area-* labels, please take a look if it's now fine.