196 regulativer / 5048 article records from 98 kommuner: one Regulativ for husholdningsaffald and one Regulativ for erhvervsaffald per kommune (husholdning: 98, erhverv: 98). 188 rows are NSTAR texts, 7 are kommune-website PDFs that replace stale NSTAR rows, and 1 (Læsø erhverv) comes from an archived kommune PDF because NSTAR cannot serve it (188 + 7 + 1 = 196). lawstatus: inforce 190, superseded 3, unverifiedcurrent 3. - NSTAR (nstar.ens.dk), Energistyrelsen's national register of municipal waste regulations. Under BEK nr. 938 af 2022 (affaldsaktørbekendtgørelsen) § 3, stk. 4, kommuner have to report adopted regulativer to NSTAR, and a regulativ is valid only once published there. The…
SAVRN Model Hub
AI Training Datasets
Each dataset with its card, its structure (every split, row count and column type), its files, its license, and the models that disclose training on it.
11,122Models
3,164Datasets
285Papers
6,792Publishers
11,852Sourced relationships
Updated 2026-10-18 · How the library is built
3,164 datasets, sorted by most downloaded.
First pass over the 30 most populous Greek municipalities (of 332), from the Diavgeia transparency portal. Target: endomorphosis/ipfsgreecemunicipallaws (new). Text-layer PDFs only (no OCR). IDs: gr-mun- -ada-; article ids -art-. The ADA (Diavgeia's unique decision number, e.g. ΨΚΧ6Ω0Ο-3ΛΝ) is transliterated 1:1 (Greek capital → one Latin letter, ΑΒΓΔΕΖΗΘΙΚΛΜΝΞΟΠΡΣΤΥΦΧΨΩ → abgdezhuiklmnjoprstyfxcw, digits kept, lower-cased) so the id is unique and reversible; the original ADA is in officialidentifier and documentnumber. Discovery: 11,306 decisions whose subject contains κανονισμός / κανονιστική / εσωτερικής υπηρεσίας (Diavgeia search/advanced.json, per municipality organisation uid, 6-month…
Bye-laws (fodhlíthe) made by Irish local authorities: county, city, and city and county councils, and their municipal / borough / metropolitan districts. Full texts are taken from the councils' own websites and split into articles. 185 bye-laws, 3,764 article records, from 23 of the 31 local authorities. English 147, Irish 38 (31 of them linked to their English original). Adoption dates run from 1997-04-28 to 2026-06-08 (62 dated, 2 of them to the month only). Schemas: MUNICIPALLAWSSCHEMA / RAWARTICLESSCHEMA from scrapers/schema.py; validatemunicipaltable returns no errors. Ids: ie-mun- - -, for example ie-mun-5-media-361-ddc14783. For a Drupal /media/N/download link the slug is media-N.…
Communal regulations (règlements communaux / Gemengereglementer) of Luxembourg's 30 most populous communes, as full texts taken from the communes' own websites, split into articles. 351 regulations, 7,972 article records, 26 communes, adoption dates 1965-02-22 … 2026-10-15. Schemas: MUNICIPALLAWSSCHEMA / RAWARTICLESSCHEMA from scrapers/schema.py; validatemunicipaltable returns no errors. Legilux (data.legilux.public.lu, SPARQL endpoint): all 34,617 acts of type RC / RTC were indexed (one query per ELI year, so no query needs an OFFSET beyond the endpoint's 10,000 limit). Legilux holds publication notices only for communal regulations: adm/rc (2016 onward: "Règlement communal – Commune de X"…
Research snapshot of Acts (ऐन) of Nepal's provincial governments from Offices of the Chief Minister and Council of Ministers (OCMCM) of Nepal's provinces. Not legal advice. The official source prevails over this corpus. Bagmati 46/47 listed (complete; 1 listed item is a duplicate PDF), Karnali 20/21 listed (complete; 1 listed item is a duplicate PDF), Gandaki pending How listed vs collected reconciles (evidence: manifests/listingindexcache.json and the instruments' source URLs): Exact duplicates dropped (same document uploaded twice; identical extracted text): np-bagmati-htsutw4 = np-bagmati-a6qm4vd, np-karnali-cnedtve = np-karnali-gmdf8w5. Failed PDF downloads: 0…
Solomon Islands laws IR release.
Research retrieval release of endomorphosis/ipfswallisfutunalaws (revision eff428b7ce17c1a960a63dce2c897635e5f0567e) packaged as country-laws-ir-graphrag/v1 (layout family skillcenter-huggingface-release/v3 / publicus-ir). Not legal advice. This is a research snapshot. The official gazette / authentic source of Wallis Futuna prevails over this corpus. Retrieved documents and graph edges are retrieval evidence only. No legal text was invented. Primary key: entrycid (CIDv1 raw sha2-256 of a canonical identity record). Integer documentindex values are compact shard pointers, not identities. Target Hub id (packaging metadata only): justicedao/ipfswallisfutunalawsir. entrycid, lawcid…
Small, browser-ready Earth observation projects for IRIS. Each project lives in its own folder and contains exactly two Cloud Optimized GeoTIFF samples. Every folder follows the same layout: The segmentation/ paths are populated by IRIS when an authenticated user saves a mask. Project data and annotation results therefore remain together in this repository. Each user needs a Hugging Face token with write access to this dataset. The newly prepared samples use Copernicus Sentinel-2 Level-2A imagery accessed through Element 84 Earth Search. See each project folder for source details and the applicable Copernicus Sentinel Data Terms and Conditions.
benchmark - evaluation - taiwan - multiple-choice - twinkle-eval - opentwbench - self-reported taskcategories: - question-answering prettyname: Jev 1.13 results on OpenTWBench sizecategories: - configname: index datafiles: - configname: formosapredictions datafiles: - configname: examsamplepredictions datafiles: path: results/jev-1.13.0/tw-exam-bench-sample/predictionsrun.jsonl Per-question predictions and summaries for TypeSafe Jev (jev-1.13.0, requested as jev-latest) on the two OpenTWBench sets: Status: self-reported. This run was not executed by the OpenTWBench maintainers and the weights are not public, so it cannot be an "Official" leaderboard row. Jev is not a chat model. It is…
The media files of the video tasks of JevArena-MM, a benchmark of JevArena decision requests whose state carries video and audio. The items, code and result tables are in the JevArena-MM repository; this dataset holds only the - A file an item references as / is at / here: the source's original file, unchanged. Items select a window of it and record its sha256. - Media whose terms forbid redistribution, or require each user to accept them first, are not hosted; JevArena-MM's fetch scripts get them from their owners or distributors (table at the end). Accept the terms on this page, log in with hf auth login (or set HFTOKEN), then in a clone of JevArena-MM: To have a file removed, open a…
key-digest-runner A minimal, self-contained deployment that runs runkeydigestresearchworkers.py continuously in Docker, driving the pydantic-researchers package (installed from PyPI) over topics drawn from keydigest.jsonl. Continuous PR reviewer (main.py) main.py watch is the autonomous merge/close loop. It polls the complete open PR queue (up to 10,000 PRs per snapshot) and reviews exactly one PR at a time. A failed PR or polling cycle is reported and
Kinex Benchmark Unified media and evidence storage for the Kinex Benchmark Space. The current index compares Codex + GPT-6 Astra / high and Kinex + GPT-6 Astra / high, with 84 selected seed-0 episodes per combination. Claude Code + Opus 5.5 / high and Kinex + Opus 5.5 / high are planned and have no results yet. Pending tasks remain pending; native verdicts, measured instructions, budgets and token counters are unchanged. Location data.json Current 164-task
l0Qwen3-1.7Bcompressionbs32n1632kt1noeosverl091seqmean One verified gzip JSONL shard per training step; 512 responses per shard. Historical compression MathVerify after thinking, without an EOS gate.
l1exactQwen3-1.7Bcompressionn100to4000alpha3e-4bs32n832k1epoch One verified gzip JSONL shard per training step; 256 responses per shard. L1-Exact: MathVerify correctness after thinking without an EOS gate, plus the original linear target-length reward. Fixed random targets 100..4000; alpha=0.0003.
l1exactQwen3-1.7Bcompressionn1000to8000alpha3e-4bs32n832k1epoch One verified gzip JSONL shard per training step; 256 responses per shard. L1-Exact: MathVerify correctness after thinking without an EOS gate, plus the original linear target-length reward. Fixed random targets 1000..8000; alpha=0.0003.
l1exactQwen3-1.7Bcompressionn400to12000bs32n832k1epoch One verified gzip JSONL shard per training step; 256 responses per shard. L1-Exact: MathVerify correctness after thinking without an EOS gate, plus the original linear target-length reward. Fixed targets 400..12000; alpha=0.000075.
Model Collections
Hand-picked starting points, each with the reason it exists.
Collection · 4 entries
Embedding models for retrieval
Sentence and document embedding models used to build retrieval systems. Dimension and sequence length matter more than size here, and both come from the publisher.
Collection · 4 entries
Models that fit on one accelerator
Models whose publisher-reported parameter count puts them within reach of a single accelerator at common precisions. Memory needed depends on precision and serving configuration, so treat the parameter count as the starting point, not the answer.
Collection · 6 entries
Open-weight text models worth knowing
Widely used open-weight language models, chosen because each one is a distinct family rather than a variant of the one above it. Selection, not a ranking.
Collection · 3 entries
Speech and audio models
Recognition and synthesis models, grouped so the two directions are easy to compare.
Related SAVRN Research
The hub sits beside SAVRN's market data and infrastructure research: what models cost to run, and what it takes to run them.
SAVRN Index
What open models cost to run
The same open-weight model priced by every host that serves it, per million tokens.
Research Hub
Data center trackers and maps
Moratoriums, permits, power, water and capital behind the facilities that run these models.
Method
How the Model Hub is built
Sources, evidence labels, refresh behaviour, and the limits of every comparison here.