GPAI Ledger › GPAI Training Transparency tracker (AI Accountability Lab (AIAL)) › Capture 12 Sep 2026

OpenLLM_France_Luciole_1_1_2026_09_08 — capture 20260912T062156Z

Filed under AI Accountability Lab (AIAL) — GPAI Training Transparency tracker, the source this project captured it from; the document itself is the filing of the model named above.

Verify: sha256sum 7698fd7f846c2e6b4b319bdaff9af1f4e4a6b79de00c74a10a8d85ff47bc34cc.pdf must equal the hash above (the filename IS the expected hash); ots verify 7698fd7f846c2e6b4b319bdaff9af1f4e4a6b79de00c74a10a8d85ff47bc34cc.pdf.20260912T062156Z.ots -f 7698fd7f846c2e6b4b319bdaff9af1f4e4a6b79de00c74a10a8d85ff47bc34cc.pdf (opentimestamps.org) proves the bytes existed no later than the attestation time — an upper bound on the capture time; the fetch time above is the archive's own record (a freshly captured proof reports 'pending' here: the calendars anchor within hours, but this archive only upgrades the stored proof to its anchor on a later run, so expect a day or two). ots verify needs a local Bitcoin Core node (a pruned one is fine); without one, ots info on the proof prints the attesting block height and merkle path to check on any block explorer.

Extracted text

Machine-extracted text (layout may be lost; the authoritative content is the stored file above).

1
TemplateforthePublicSummaryofTraining
ContentforGeneral-PurposeAImodels
VersionoftheSummary :V1.1,firstpublishedversion.Nopreviouslypublished
versionofthisSummaryexists.
Lastupdate:31/07/2026
Generalinformation
1.Generalinformation
1.1.Provideridentification
Providernameandcontactdetails
LINAGORA—SIREN431473669,registeredundernumber431473669R.C.S.
Nanterre.
Registeredoffice:VillaGoodTech,37ruePierrePoli,92130Issy-les-Moulineaux,
France.
ContactformattersrelatingtothisSummary:jplorre@linagora.com
Websites:
 https://www.linagora.com
 https://openllm-france.fr
LINAGORAisasmallormedium-sizedenterprisewithinthemeaningofCommission
Recommendation2003/361/EC.TheSMEthresholdreferredtoinSection2.3ofthis
templateisthereforetheapplicableone.
Authorisedrepresentativenameandcontactdetails
Notapplicable.TheproviderisestablishedintheUnion(France);Article54AIActdoes
notapply.
1.2.Modelidentification
Versionedmodelname(s)
This Summary covers the following models, whose training content is identical, in
accordancewithpoint30oftheCommissionExplanatoryNotice:
 Luciole-1B-Instruct-1.1:
https://huggingface.co/OpenLLM-France/Luciole-1B-Instruct-1.1
 Luciole-8B-Instruct-1.1:
https://huggingface.co/OpenLLM-France/Luciole-8B-Instruct-1.1
 Luciole-23B-Instruct-1.1:
https://huggingface.co/OpenLLM-France/Luciole-23B-Instruct-1.1
AllmodelsarereleasedundertheApache2.0licencewithpubliclyavailableweights,
includingintermediatecheckpointsavailablehere:
https://dl.labs.linagora.com/files/models/OpenLLM-France/
2
Modeldependencies
The Luciole Instruct 1.1 models are post-trained versions of the Luciole base models:
Luciole-1B-Base,Luciole-8B-Base,Luciole-23B-Basefoundhere:
https://huggingface.co/collections/OpenLLM-France/luciole-llm. Trainingtookplacein
threestages,witheachstagestartingfromthelastcheckpointofthepreviousphase:

Luciole-xB-SFT-Thinking: a supervised fine-tuning of the corresponding base
modelusingdatawiththinkingtraces.

Luciole-xB-SFT-Instruct: a supervised fine-tuning of the corresponding base
modelusingdatawithoutthinkingtraces.

Luciole-xB-Instruct:alignmentoftheSFT-InstructmodelusingDPO.
DateofplacementofthemodelontheUnionmarket
09July2026
1.3Modalities,overalltrainingdatasizeandother
characteristics
Modality
☒Text
Textistheonlymodalitypresentinthetrainingdata
Trainingdatasize
☒Lessthan1billiontokens
Approximately4BforSFTThinking,2.3BforSFT,0.75BforSFTDPOafterpre-
processing
Typesofcontent
SFT:Examplesofinstructionswithresponsesinvariousdomainsincluding:code,
math,STEM,chat,RAG,toolcalling,translationandNLI.
DPO:Examplesofinstructionswithpairsofresponseswhereoneislabelled“chosen”
andtheother“rejected”.SamedomainsasSFT,excludingtranslationandincluding
safety
Latestdateofdataacquisition/collectionformodeltraining
June2026
Descriptionofthelinguisticcharacteristicsoftheoveralltrainingdata
TheLuciole-PostTraining-Datasetv1.1consistsalmostexclusivelyofEnglishlanguage
data,withtheexceptionofsomemultilingualrepresentationintheNemotron
datasetsandsafetydatasets.
Otherrelevantcharacteristicsoftheoveralltrainingdata
Thecategoriesofdatausedduringtrainingbreakdownroughlyasfollows:
 Chat:25%
 Math:13%
 STEM:14%
 Code:16%
 Toolcalling:11%
 NLI:13%
3
 Safety:5%
 Other(includingmultilingual):3%
Additionalcomments
Tokeniservocabulary:128000tokens.
Contextlength:16,384tokens.
Training phases: SFT Thinking used 2.6 million samples, SFT Instruct used 2.1 million,
andInstructDPOused200,000samples.
Training infrastructure: training was carried out on NVIDIA H100 80 GB GPUs of the
JeanZaysupercomputer(GENCI/IDRIS).
Funding: development was carried out by LINAGORA within the OpenLLM-France
consortiumwithfundingfromBpifranceundertheFrance2030programme.
2.Listofdatasources
2.1.Publiclyavailabledatasets
Haveyouusedpubliclyavailabledatasetstotrainthemodel?
☒Yes
Ifyes,specifythemodality(ies)ofthecontentcoveredbythedatasets
concerned
☒Text
Listoflargepubliclyavailabledatasets
Muchofthetrainingcontentforthesupervisedfine-tuning(SFT)phasesofInstruct1.1
models consists of pre-packaged, openly licensed datasets compiled by third parties
orbyOpenLLMpartners.Ratherthanapplythe3%materialitythreshold,theprovider
disclosesthecompletelistdatasetsused.Thatlistispublishedandkeptuptodateat:
https://huggingface.co/datasets/OpenLLM-France/Luciole-PostTraining-Dataset-1.1
ForSFTThinking,weusedthefollowingthird-partydatasets:
 DOLCI [Think Persona Precise IF, Think Precise IF, SYNTHETIC-2-SFT-Verified]
(ODC-BY)
 NemotronPosttrainingv3[ChatEnglish,Mathv2,ScienceMCQ] (CC-BY-4.0)
 NemotronPosttrainingv2[Code,Mathfr](CC-BY-4.0)
 OpenCodeReasoning(CC-BY-4.0)
 Nemotronagenticsftv2(Apache-2.0)
 smolagenttoolcalling(Apache-2.0)
 PleiasRAG(CC-BY-4.0)
ForSFTwithoutthinking:
 DOLCI[ThinkPreciseIF,Sciriff,PythonAlgos,Flan2,Logicpuzzles] (ODC-BY)
 NemotronPosttrainingv3[ChatEnglish](CC-BY-4.0)
 NemotronPosttrainingv2[Code,Stem](CC-BY-4.0)
 PleiasRAG(CC-BY-4.0)
 Linagora[PersonasMath,HotpotQA,TatQA,Hardcoded](CC-BY-4.0)
 XLAM(CC-BY-4.0)
 Hermes(Apache-2.0)
4
 When2call(CC-BY-4.0)
 Paradocs(Apache-2.0)
 Croissant(CC-BY-SA-4.0)
 Smol[instruct,rewrite,summarize](Apache-2.0)
ForDPO,weusedpromptsfromthefollowingdatasets:
 DOLCI [Think Persona Precise IF, Think Precise IF, Sciriff, Python Algos, Flan2 ]
(ODC-BY)
 NemotronPosttrainingv3[ChatEnglish,Mathv2,ScienceMCQ] (CC-BY-4.0)
 NemotronPosttrainingv2[Code,Stem](CC-BY-4.0)
 NemotronSafety(CC-BY-4.0)
 OpenMathInstruct(NVIDIALicense)
 PleiasRAG(CC-BY-4.0)
 Smol[instruct,rewrite,summarize](Apache-2.0)
 When2call(CC-BY-4.0)
 XLAM(CC-BY-4.0)
Topreprocessthesplitdatasets,wecheckedforthepresenceofnamesofLLMsand
companies(Claude,Amazon,etc.)aswellasforChineseandRussian.Giventhatmost
ofthedatawasgeneratedwithopensourcemodelsfromQwenandDeepSeek,there
wasapreponderanceofChineseandRussianscriptintermingledwithourlanguages
ofinterest.Samplescontaininganyofthesewereremovedentirely.Theprepocessing
scriptscanbefoundinthisfolderoftheLuciole-Trainingrepository:
https://github.com/OpenLLM-France/Luciole-
Training/tree/main/data/processing/posttraining
Generaldescriptionofotherpubliclyavailabledatasetsnotlistedabove
None. All publicly available datasets used are enumerated in the published corpus
documentationreferredtoabove.
Thecorpusisrestrictedtomaterialdistributedunderopenlicences.
Additionalcomments
ThecorpusitselfispublishedunderCCBY-SA4.0.Theprocessingscriptsarepublicat
https://github.com/OpenLLM-France/Luciole-
Training/tree/main/data/processing/posttraining
VerificationofthestatementsinthisSummaryisthereforepossibleatsourcelevel.
List of data sources
2.2Privatenon-publiclyavailabledatasetsobtainedfromthird
parties
2.2.1.Datasetscommerciallylicensedbyrightsholdersortheir
representatives
Haveyouconcludedtransactionalcommerciallicensingagreement(s)with
rightsholder(s)orwiththeirrepresentatives?
☒No
5
2.2.2.Privatedatasetsobtainedfromotherthirdparties
Haveyouobtainedprivatedatasetsfromthirdpartiesthatarenotlicensedas
describedinSection2.2.1,suchasdataobtainedfromprovidersofprivatedatabases,
ordataintermediaries?
☒No
2.3Datacrawledandscrapedfromonlinesources
Werecrawlersusedbytheprovideroronbehalfof?
☒No
2.4Userdata
WasdatafromuserinteractionswiththeAImodel(e.g.userinputandprompts)used
totrainthemodel?
☒No
Wasdatacollectedfromuserinteractionswiththeprovider’sotherservicesor
productsusedtotrainthemodel?
☒No
Nodatafromusersoftheprovider'sservicesorproductswasusedtotrainthe
models.Inparticular,noLinTO,TwakeorotherLINAGORAproductdata,andnologs
frompublicLucioledemonstrators,enteredthepre-trainingorpost-trainingcorpora.
2.5Syntheticdata
WassyntheticAI-generateddatacreatedbytheproviderorontheirbehalfto
trainthemodel?
☒Yes
Ifyes,modalityofthesyntheticdata
☒Text
Ifyes,specifythegeneral-purposeAImodel(s)usedtogeneratethesynthetic
dataifavailableonthemarket
GenerationwasusedintheDPOphase.Withtheexceptionofdataforsafety
alignment,thealignmentpairsweregeneratedsyntheticallywithadeltalearning
approach:allpairsweregeneratedwithQwen3-32BandQwen3-0.6Bandtheformer
werelabelledastheacceptedresponses.Safetydataweregeneratedwithamixture
ofQwen3-14B,Ministral-3-14B-Instruct,andaninterimcheckpointofLuciole-8B-
Instruct,afterSFT.PairswerejudgedwithbothMinistral-14B-ReasoningandQwen3-
14B.Apairwasincludedonlyifbothmodelsagreedontheirlabels.
InformationaboutotherAImodels,includingprovider’sownAImodel(s)not
availableonthemarket,usedtogeneratesyntheticdatatotrainthemodelto
whichthisSummaryapplies
6
AnaninterimcheckpointofLuciole-8B-Instruct,afterSFT,wasusedtomakesome
safetydata.Thetrainingdatawasthesameasdescribedabove.
Additionalcomments
Several third-party post-training datasets used by the provider were themselves
produced by AI models. Because they were obtained as publicly available datasets
ratherthangeneratedbyoronbehalfoftheprovider,theyarereportedunderSection
2.1.
Beforeuse,theproviderremovedreferencestothird-partymodelnamesandcompany
names, and removed Chinese and Russian script, in order to limit the transfer of
provenanceartefactsandculturalbias.
2.6Othersourcesofdata
HavedatasourcesotherthanthosedescribedinSections2.1to2.5beenusedto
trainthemodel?
☒No
AnnotationofpreferencepairsforDPOwasdeterminedautomaticallybyusinga
largermodeltoproducetheacceptedresponsesandaverysmallmodeltogenerate
therejectedresponses.
3.Dataprocessingaspects
3.1.Respectofreservationofrightsfromtextanddatamining
exceptionorlimitation
AreyouaSignatorytotheCodeofPracticeforgeneral-purposeAImodelsthat
includescommitmentstorespectreservationsofrightsfromtheTDMexception
orlimitation?
☒Yes
Describe the measures implemented before model training to respect
reservations of rights from the TDM exception or limitation before and during
data collection, including the opt-out protocols and solutions honoured by the
provider or, as applicable, by third parties from which datasets have been
obtained:
LINAGORAisasignatorytotheGeneral-PurposeAICodeofPracticeandadherestoits
Copyright chapter, including Measure 1.3 on the identification of and compliance with
reservationsofrights.
Thiscommitmentismaintainedthroughtestrictionofthecorpustoopenlylicensed
material.Thecorpuswasassembledexclusivelyfromsourcesdistributedunderopen
licencesorinthepublicdomain.Contentwhoserightsholdershadreservedrights
underArticle4(3)ofDirective(EU)2019/790wasthereforenotatargetofcollection.
BecausethedatacomesentirelyfromsyntheticgenerationorstandardNLPdatasets
7
(e.g.,HotPot,FLAN),noparticularmeasuresweretakentofilterprotectedworkor
personaldata(asthisisnotexpectedtobeaproblem).
Fordatasetsobtainedfromthirdparties,theprovideradditionallyreliesonthe
collectionpracticesandrights-reservationcomplianceoftheupstreamcompilers,as
documentedbythosecompilers.
3.2Removalofillegalcontent
Generaldescriptionofmeasurestaken
Dataiseithergeneratedusingopenmodelsorchosenfromopendatasetspublished
bythirdpartiesafterextensivefilteringorcarefulurlselectiontotargetspecific
domainssuchasmath,codeandscience.Itistakennotfrommassive,uncontrolled
webcontent,thusillegalcontentisnotexpectedtobeaproblem.Highlytoxicor
dangerouscontentisnotexpectedtobepresenteitherexceptinthecaseofsafety
datausedintheDPOphase.Thiscontentcannotbeheavilyfilteredduetoitspurpose
intrainingmodelswhatnottosay.
1. D
3.3.Otherinformation
Otherrelevantinformationaboutdataprocessing
Topreprocessthesplitdatasets,wecheckedforthepresenceofnamesofLLMsand
companies(Claude,Amazon,etc.)aswellasforChineseandRussian.Giventhatmost
ofthedatawasgeneratedwithopensourcemodelsfromQwenandDeepSeek,there
wasapreponderanceofChineseandRussianscriptintermingledwithourlanguages
ofinterest.Samplescontaininganyofthesewereremovedentirely.Theprepocessing
scriptscanbefoundinthisfolderoftheLuciole-Trainingrepository:
https://github.com/OpenLLM-France/Luciole-
Training/tree/main/data/processing/posttraining