ai. ada -se ice.eu
FIZ p o ides a sel hos ed AI in as uc u e
con aining Fai -Way Se ice und RADAR Keywo d
Se ice.
•20 GB GPU
•Focus on SLM and domain-speci ic AI models
(e.g. o se ices o NFDI communi ies, i.e.
RADAR4Chem, RADAR4Cul u e,
RADAR4Memo y)
•Da a p i acy
•Calculable cos s
FAIRness Assessmen
Fai -Way Se ice
•SLM/LLM-powe ed FAIR assessmen
•Open Sou ce (Gi Hub)
•Suppo o JSON, XML, HTML, …
•Scalable
•REST API / Py hon Fas API
•Docke Compose
•Ollama o model loading
© 2024 Copy igh : Anmol
Sha ma, Da abases and
In o ma ion Sys ems (i5), RWTH
Aachen
Au ho s
S e an.Ho mann@ iz-ka ls uhe.de
Ke s in.Sol au@ iz-ka ls uhe.de
Felix.Bach@ iz-ka ls uhe.de
Licensed unde CC-BY 4.0 | h ps://c ea i ecommons.o g/licenses/by/4.0 (Sep embe 2025)
Enhancing FAIR Resea ch Da a
Managemen wi h AI Suppo
RADAR is an es ablished
in e disciplina y eposi o y o he
a chi ing, publica ion, and long-
e m p ese a ion o esea ch
da a.
I is de eloped and ope a ed by
FIZ Ka ls uhe.
We a e cu en ly explo ing how
AI can enhance me ada a quali y
h ough ea u es such as
au oma ed me ada a ex ac ion
and FAIRness e alua ion.
TS4NFDI
Te minology
Se ices
Inpu / In e ence da a
Me ada a
Ti le(s), Desc ip ion(s), Au ho (s), Con ibu o (s), Funde (s),
Rela ed Iden i ie (s), Loca ion …
File Con en s
…
…
ai. ada -se ice.eu
Gene ic LLM like Cha GPT
Cha AI (GWDG)
E c.
AI
RADAR
Keywo d Se ices Fai -Way
Se ice
Da ase Lis ing
RADAR Keywo d
Se ice
•SLM/LLM-powe ed keywo d sea ch
•Based on KeyBERT API / KeyLLM: REST API
/ Py hon Fas API
•Can be connec ed wi h ei he
Cha GPT o domain-speci ic SLM
•Seman ic Suppo ia TS4NFDI
© 2024 Copy igh : FIZ Ka ls uhe
Easy o implemen ; lexible use o LLM
(Cha GPT, Mis al).
High cos s; non- ep oducible esul s;
misleading sco es (e.g. nea ly emp y s.
ex ended me ada a); p i acy conce ns
wi h ex e nal se ices.
Open Sou ce; uns on own
in as uc u e; s anda dized me ics;
high quali y epo s; compa ible also
wi h any domain-speci ic me ada a;
p omising es esul s
We ocus on Me hod 2 due o p omising esul s, educed p i acy
conce ns, and he Fai -Way epo , ensu ing as e implemen a ion.
No ye all FAIRsFAIR checks
implemen ed; AI may in oduce e o s;
high CPU load on low-end ha dwa e;
equi es dedica ed se ice.
LLM & P omp Enginee ing.
LLM calcula es a FAIRness sco e
(0–100%) based on gi en me ada a.
Me hod 1
Fai -Way Se ice.
Re u ns FAIRness sco es wi h de ailed
eedback (FAIRsFAIR Da a Objec
Assessmen Me ics 0.5)
Me hod 2
Aim: Mo i a e use s o enhance a da ase wi h FAIR and signi ican
(seman ic) me ada a.
Conclusion
Me ada a Enhancemen
Easy o implemen ; lexible
(gene ic o domain-speci ic
models); analyzes bo h me ada a
and ile con en .
Quali y depends on model and
p omp design, non- ep oducible
esul s, high cos s; p i acy
conce ns; poo XML me ada a
ou pu .
Domain-speci ic en ichmen wi h
seman ic p ecision; s anda dized
e minology imp o es
in e ope abili y; KeyBERT is
OpenSou ce and hus anspa en
how i wo ks.
We ocus on Me hod 2 due o p omising esul s,
plannable cos s and p i acy conce ns.
Me ada a enhancemen is limi ed
o keywo ds; depends on
co e age and quali y o
e minologies, equi es a
dedica ed se ice.
LLM & P omp Enginee ing.
A LLM like Cha GPT enhances RADAR
me ada a. To ge be e esul s, se e al
hin s and XML examples a e p o ided
in he p omp .
Me hod 1
RADAR Keywo d Se ice.
Keywo ds a e ex ac ed om me ada a
and ile con en ia KeyBERT. These
keywo ds a e mapped o seman ic
e ms i possible.
Me hod 2
Aim: Suppo use s in en iching da ase s/ iles wi h signi ican
me ada a based on me ada a, ile con en and ex e nal esou ces.
Conclusion
Mo e In o ma ion
in o@ ada -se ice.eu
www. ada -se ice.eu
…
