FAIR for machine learning
FAIR was written for data, extended to software (FAIR4RS -
rseng-fair-software), and is now being worked out for machine
learning (the RDA FAIR4ML interest group leads the metadata
standardization). ML needs its own treatment because a model is
neither data nor ordinary software: its behavior is inseparable
from its training data, its "source" includes weights that no
license default covers, and reuse without documentation reproduces
biases invisibly. The practical FAIR-ML unit is the CLUSTER -
model + data + code + evaluation + paper - linked both ways with
persistent identifiers.
Findable: metadata that machines and reviewers read
- Model cards are the model's README and its FAIR metadata in one:
intended use and out-of-scope uses, training data description,
evaluation results with conditions, limitations and biases,
licensing. Hub-hosted cards (Hugging Face's structured format)
double as searchable metadata; write them from the actual
training run, not memory (the experiment config and logs are the
source - rseng-reproducibility).
- Datasets get Croissant: the MLCommons format describes ML
datasets (schema, distribution, provenance) in machine-readable
form that major hubs and search index - the ML-ready complement
to the generic dataset documentation in rseng-data-management.
- Identifiers: archive released models and dataset snapshots with
DOIs (rseng-archiving; hub storage is not preservation), and
cross-link model card, dataset record, code repository
(rseng-citation-metadata) and paper so each cites the others.
Accessible and licensed honestly
- State exactly what is released: weights, architecture, training
code, data, all or some - and license each part explicitly.
Code licenses do not fit weights cleanly; dedicated model
licenses (including use-restricted ones) exist, and "weights on
a hub with no license" is the ML version of unlicensed code:
unusable (rseng-licensing, rseng-license-compliance for
compatibility when models build on models).
- Gated access is legitimate (safety, privacy, data terms) when
the CONDITIONS are stated and the metadata stays public - as
closed as necessary, as open as possible
(rseng-regulatory-compliance where personal data trained the
model; the EU AI Act documentation duties overlap helpfully
with a good model card).
Interoperable: formats and conventions
- Export to exchange formats where the ecosystem has them (ONNX-
class interchange for deployment; standard checkpoint formats
per framework) and pin the framework versions that load the
weights (rseng-reproducible-environments) - a checkpoint nobody
can load is a dead artifact.
- Follow the target hub's metadata conventions (tags, task
taxonomy, evaluation fields) - interoperability for models is
largely hub-convention compliance today, while FAIR4ML
standardization matures.
Reusable: the documentation that prevents misuse
Reuse-readiness is a checklist an agent can run on any model
(the user's own before release, third-party models before
adoption - the rseng-software-reuse evaluation instinct):
- Can I tell what data trained it, and under what terms?
- Are evaluation claims reproducible - metric, dataset version,
split, conditions stated (rseng-research-integrity for the
numbers)?
- Are limitations and known failure modes documented, or must I
rediscover them?
- Is the license compatible with my use, including the data's
terms flowing through?
- Can I load it - format, framework versions, hardware
requirements (rseng-gpu-computing)?
A "no" on any of these is a finding: fix it for your own models,
weigh it for others'. Record AI/ML provenance in aidecl.yaml
(rseng-ai-declaration) - a project that trains models with AI
assistance has two provenance layers, and both belong in the
record.
Working with this skill
This skill is source-independent: its authority is the FAIR4ML
work, the Croissant specification and model-card practice linked
below. It extends rseng-fair-software to ML artifacts; rseng-fairguard
assesses the software side.
Learn more (verified):
Related skills
Check whether any of these applies before moving on:
- rseng-archiving - DOIs for model and dataset snapshots
- rseng-data-management - dataset documentation beneath Croissant
- rseng-fair-software - the general FAIR baseline
- rseng-gpu-computing - hardware requirements for model reuse
- rseng-licensing - licensing weights and training data
- rseng-reproducibility - model cards written from actual runs
1---2name: rseng-fair-ml3description: Covers applying FAIR principles to machine learning artifacts: making models findable and reusable with model cards and rich repository metadata, documenting datasets with Croissant and datasheet-style records, licensing models and weights, linking the model-data-code-paper cluster with persistent identifiers, and the RDA FAIR4ML metadata direction. Use when a project trains, fine-tunes, publishes or reuses ML models or ML-ready datasets, when the user mentions model cards, Croissant, datasheets, FAIR4ML or model licensing, when a model heads to a hub or archive, or when evaluating whether a third-party model is documented well enough to build on. (General software FAIR is rseng-fair-software; automated repository scoring is rseng-fairguard.)4license: CC-BY-4.05---67# FAIR for machine learning89FAIR was written for data, extended to software (FAIR4RS -10rseng-fair-software), and is now being worked out for machine11learning (the RDA FAIR4ML interest group leads the metadata12standardization). ML needs its own treatment because a model is13neither data nor ordinary software: its behavior is inseparable14from its training data, its "source" includes weights that no15license default covers, and reuse without documentation reproduces16biases invisibly. The practical FAIR-ML unit is the CLUSTER -17model + data + code + evaluation + paper - linked both ways with18persistent identifiers.1920## Findable: metadata that machines and reviewers read2122- Model cards are the model's README and its FAIR metadata in one:23 intended use and out-of-scope uses, training data description,24 evaluation results with conditions, limitations and biases,25 licensing. Hub-hosted cards (Hugging Face's structured format)26 double as searchable metadata; write them from the actual27 training run, not memory (the experiment config and logs are the28 source - rseng-reproducibility).29- Datasets get Croissant: the MLCommons format describes ML30 datasets (schema, distribution, provenance) in machine-readable31 form that major hubs and search index - the ML-ready complement32 to the generic dataset documentation in rseng-data-management.33- Identifiers: archive released models and dataset snapshots with34 DOIs (rseng-archiving; hub storage is not preservation), and35 cross-link model card, dataset record, code repository36 (rseng-citation-metadata) and paper so each cites the others.3738## Accessible and licensed honestly3940- State exactly what is released: weights, architecture, training41 code, data, all or some - and license each part explicitly.42 Code licenses do not fit weights cleanly; dedicated model43 licenses (including use-restricted ones) exist, and "weights on44 a hub with no license" is the ML version of unlicensed code:45 unusable (rseng-licensing, rseng-license-compliance for46 compatibility when models build on models).47- Gated access is legitimate (safety, privacy, data terms) when48 the CONDITIONS are stated and the metadata stays public - as49 closed as necessary, as open as possible50 (rseng-regulatory-compliance where personal data trained the51 model; the EU AI Act documentation duties overlap helpfully52 with a good model card).5354## Interoperable: formats and conventions5556- Export to exchange formats where the ecosystem has them (ONNX-57 class interchange for deployment; standard checkpoint formats58 per framework) and pin the framework versions that load the59 weights (rseng-reproducible-environments) - a checkpoint nobody60 can load is a dead artifact.61- Follow the target hub's metadata conventions (tags, task62 taxonomy, evaluation fields) - interoperability for models is63 largely hub-convention compliance today, while FAIR4ML64 standardization matures.6566## Reusable: the documentation that prevents misuse6768Reuse-readiness is a checklist an agent can run on any model69(the user's own before release, third-party models before70adoption - the rseng-software-reuse evaluation instinct):71721. Can I tell what data trained it, and under what terms?732. Are evaluation claims reproducible - metric, dataset version,74 split, conditions stated (rseng-research-integrity for the75 numbers)?763. Are limitations and known failure modes documented, or must I77 rediscover them?784. Is the license compatible with my use, including the data's79 terms flowing through?805. Can I load it - format, framework versions, hardware81 requirements (rseng-gpu-computing)?8283A "no" on any of these is a finding: fix it for your own models,84weigh it for others'. Record AI/ML provenance in aidecl.yaml85(rseng-ai-declaration) - a project that trains models with AI86assistance has two provenance layers, and both belong in the87record.8889## Working with this skill9091This skill is source-independent: its authority is the FAIR4ML92work, the Croissant specification and model-card practice linked93below. It extends rseng-fair-software to ML artifacts; rseng-fairguard94assesses the software side.9596Learn more (verified):97 - https://www.rd-alliance.org/groups/fair-machine-learning-fair4ml-ig/ -98 RDA FAIR4ML interest group99 - https://mlcommons.org/working-groups/data/croissant/ -100 Croissant ML dataset format101 - https://huggingface.co/docs/hub/en/model-cards - model card102 format and guidance103 - https://www.gofair.foundation/fair-principles - the FAIR principles104 - https://onnx.ai - ONNX model interchange format105106<!-- related-skills:begin -->107108## Related skills109110Check whether any of these applies before moving on:111112- rseng-archiving - DOIs for model and dataset snapshots113- rseng-data-management - dataset documentation beneath Croissant114- rseng-fair-software - the general FAIR baseline115- rseng-gpu-computing - hardware requirements for model reuse116- rseng-licensing - licensing weights and training data117- rseng-reproducibility - model cards written from actual runs118119<!-- related-skills:end -->
fdiblen