iiinova — where your digital self keeps evolving

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 22/10/2026

I think there is quite a difference in how you do it. I really like the notions of answerability. I would love to interview of authors of some papers I review. Because I am afraid that the author that will be able to answer critical questions. link.springer.com/article/10.1...

link.springer.com

AI and the Future of Publishing: Not Detection, but Answerability - Philosophy & Technology

Publishing has addressed the challenges posed by large language models (LLMs) through a provenance strategy: detectors, declarations, watermarks, and attestations. Provenance matters, but as a test of...

110 1

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 21/10/2026

Swedish carrying the Scandinavian languages — feels like there is meme potential here

110 1

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 21/10/2026

If you are looking for permissible data, take a look at the dynawords (also comes in Polish and Dutch)

040 4

Reposted by Kenneth Enevoldsen

Vladimir Salnikov @v4ldelund.bsky.social · 16/10/2026

Norwegian Dynaword just crossed 10B tokens I am very happy to see open source data growth for all Scandinavian languages and it makes me definitely more optimistic about future of training models on ethically collected permissible data

1203 20

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 13/10/2026

Amazing for on this @v4ldelund.bsky.social - I really think a lot of people in digital heritage will love this!

010 1

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 07/10/2026

Submitted my two accepted posters for EMNLP today, if you are there an want meet up feel free to reach out! I also have a not so @mteb.org themed poster for a workshop on socio-cultural alignment -- stay tuned for that one

Poster on MVEB: Massive Video Embedding Benchmark

Poster on the paper: One Prompt is Not Enough: Instruction sensitivity Endermines Embedding Model Evaluation.

0102 10

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 05/10/2026

paper, code (generation code, analysis,...), data (templates, evaluation logs,...) everything is public, even the annotation process is in git commits. You can find it here: huggingface.co/papers/2610....

huggingface.co

Paper page - Multilingual GSM-Symbolic: What determines capability transfer across languages?

Join the discussion on this paper page

020 2

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 05/10/2026

Very happy to announce our new paper "MULTILINGUAL GSM-SYMBOLIC: What determines capability transfer across languages?" Where we construct explore how capabilities transfer between languages and show how these insights impact both evaluation and development.

152 5

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 29/09/2026

1B is a pretty high number — I think the fact that we are already beating common crawl is quite impressive

010 1

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 29/09/2026

Kinda what mteb does with its warning if the authors didn’t specify training data

111 1

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 24/09/2026

Sure, mostly legitimate PRs written by users using AI, everything that has the label ‘long-pr-desc’. You see in the comment that gets autogenerated that we focus most on maintainer time (rather than AI usage) github.com/embeddings-b...

github.com

Pull requests · embeddings-benchmark/mteb

MTEB: State-of-the-art evaluation of embeddings across languages and modalities - Pull requests · embeddings-benchmark/mteb

120 2

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 24/09/2026

I would say it depends on how big of a problem it is. We (mteb) try to respect valid bug reports - since whatever you do to the model response also encourage or discourage contributions by legitimate contributors But we have made some efforts especially on PRs to discorage overly long messages

110 1

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 24/09/2026

General purpose classifier don’t sell as well (sadly)

110 1

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 23/09/2026

We do find that blame has risen notably in recent years, but also that who blames is highly determined by opposition status, a phenomenon we name 'political contrasting'.

120 2

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 23/09/2026

We often say that the political environment has grown more hostile in recent years, but has it? Markus Lundsfryd Jensen, Rune Egeskov Trust, Sara Kolding, and I examine this in a recent work exploring the dynamics of blame attribution in the Danish Parliament from 1997-2026.

194 9

Reposted by Kenneth Enevoldsen

MTEB @mteb.org · 03/09/2026

New on MTEB: google/gemini-embedding-2, which takes interleaved text, image, audio and video in one input. Its first results are on ViDoRe(v2), visual document retrieval: 62.37, first of the 13 single-vector models with complete results. Model and results by whybe-choi.

New model on MTEB: google/gemini-embedding-2 — First of the 13 single-vector models on ViDoRe(v2). Bar chart of 13 of the 13 models with complete results on ViDoRe(v2) (4 tasks), each row given with its rank on that board. 1. google/gemini-embedding-2, 62.37, size not published (this release); 2. nanovdr/NanoVDR-S-Multi, 61.93, 69M parameters; 3. jinaai/jina-embeddings-v4, 58.23, 3.9B parameters; 4. eagerworks/eager-embed-v1, 57.20, 4.4B parameters; 5. ibm-granite/granite-vision-3.3-2b-embedding, 56.92, 3.0B parameters; 6. voyageai/voyage-multimodal-3, 54.94, size not published; 7. royokong/e5-v, 46.55, 8.4B parameters; 8. google/siglip-so400m-patch14-384, 33.00, 878M parameters; 9. TIGER-Lab/VLM2Vec-Full, 31.82, 4.1B parameters; 10. jinaai/jina-clip-v2, 26.94, 865M parameters; 11. laion/CLIP-ViT-bigG-14-laion2B-39B-b160k, 19.35, 2.5B parameters; 12. openai/clip-vit-base-patch16, 11.26, 150M parameters; 13. jinaai/jina-clip-v1, 9.74, 223M parameters. Model and results by whybe-choi. Source: MTEB, leaderboard.mteb.org.

111 1

Reposted by Kenneth Enevoldsen

MTEB @mteb.org · 01/09/2026

New benchmark: MTEB(slk, v1) is MTEB's first Slovak board — 31 tasks across 7 task types, 17 built for it: parliamentary sentiment, fact-checking NLI, pharmacy Q&A, news clustering. Of the 8 models scored on all 31, intfloat/multilingual-e5-large-instruct leads. Contributed by andrejridzik.

New benchmark on MTEB: MTEB(slk, v1) — 31 tasks across 7 task types — 17 of them built for this board. Stacked column chart, tasks per type, new tasks above existing ones: Classification 7 (3 new), BitextMining 6 (1 new), Clustering 5 (4 new), Retrieval 5 (1 new), Reranking 3 (3 new), PairClassification 3 (3 new), STS 2 (2 new). 8 models have complete 31-task coverage so far. Six of the 17 new tasks: SlovakPharmacyDrMaxReranking (Reranking) — Pharmacists’ answers to medication and health questions, from a Slovak pharmacy site; DemagogSKNLI (PairClassification) — Fact-checkers’ evidence paired with the political claims it supports or refutes; SlovakParlaSentClassification (Classification) — Sentiment in Slovak parliamentary debate, three levels; PravdaSKTagClustering (Clustering) — News articles from Pravda.sk, grouped by their editorial tags; SMESumRetrieval (Retrieval) — Article introductions as queries over 80,000 SME news documents; SlovakSTS (STS) — A professional Slovak translation of the STS Benchmark. Benchmark by andrejridzik. Source: MTEB, leaderboard.mteb.org.

041 4

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 26/08/2026

Hi Trev, can you explain the difference between direct infringement and secondary infringement?

000 0

Reposted by Kenneth Enevoldsen

MTEB @mteb.org · 20/08/2026

New on MTEB: malteos/most-embed-de, a German retrieval model fine-tuned from Nemotron-3-Embed-1B, takes 1st of 106 models with full coverage on a custom 6-task German-retrieval subset at 58.07 — beating the runner-up at roughly 1/15th the active parameters.

New model on MTEB: malteos/most-embed-de — Highest score on German retrieval among 106 full-coverage models — 15× smaller than the runner-up. Scatter plot of German Retrieval (6 tasks, custom subset) against Active parameters for 104 German-retrieval models with complete 6-task coverage. malteos/most-embed-de scores 58.07 at 872M active parameters. Compared against codefuse-ai/F2LLM-v2-14B at 13.214B, 55.73; codefuse-ai/F2LLM-v2-8B at 6.947B, 55.22; Snowflake/snowflake-arctic-embed-l-v2.0 at 312M, 54.24; codefuse-ai/F2LLM-v2-4B at 3.634B, 53.98; codefuse-ai/F2LLM-v2-1.7B at 1.41B, 52.91; Snowflake/snowflake-arctic-embed-m-v2.0 at 113M, 51.66; hotchpotch/bekko-embedding-v1-a25m at 25M, 51.26; jinaai/jina-embeddings-v3 at 316M, 49.22; hotchpotch/bekko-embedding-v1-a8m at 8M, 48.90; codefuse-ai/F2LLM-v2-0.6B at 441M, 48.82; codefuse-ai/F2LLM-v2-330M at 198M, 48.79; GritLM/GritLM-7B at 7.109B, 47.68; intfloat/e5-mistral-7b-instruct at 6.98B, 46.69; intfloat/multilingual-e5-large-instruct at 304M, 45.96; OrdalieTech/Solon-embeddings-large-0.1 at 304M, 45.82; intfloat/multilingual-e5-large at 304M, 44.91; mteb/baseline-bm25s at 0 (static), 44.80; Lajavaness/bilingual-embedding-large at 304M, 44.71; intfloat/multilingual-e5-base at 86M, 43.43; HIT-TMG/KaLM-embedding-multilingual-mini-v1 at 358M, 43.39; codefuse-ai/F2LLM-v2-160M at 62M, 43.05; Lajavaness/bilingual-embedding-base at 86M, 42.94; intfloat/multilingual-e5-small at 22M, 42.62; aari1995/German_Semantic_STS_V2 at 304M, 40.99; ibm-granite/granite-embedding-278m-multilingual at 86M, 40.25; Lajavaness/bilingual-embedding-small at 22M, 40.17; HIT-TMG/KaLM-embedding-multilingual-mini-instruct-v1 at 358M, 39.91; codefuse-ai/F2LLM-v2-80M at 32M, 39.21; ibm-granite/granite-embedding-107m-multilingual at 11M, 37.91; jinaai/jina-embedding-b-en-v1 at 85M, 36.53; manu/sentence_croissant_alpha_v0.2 at 1.214B, 36.22; intfloat/e5-large-v2 at 304M, 35.97; sentence-transformers/static-similarity-mrl-multilingual-v1 at 0 (static), 35.55; intfloat/e5-large at 304M, 33.95; Snowflake/snowflake-arctic-embed-l at 304M, 33.86; thenlper/gte-large at 304M, 33.82; sentence-transformers/paraphrase-multilingual-mpnet-base-v2 at 86M, 33.75; WhereIsAI/UAE-Large-V1 at 304M, 33.33; BAAI/bge-large-en-v1.5 at 304M, 33.24; mixedbread-ai/mxbai-embed-large-v1 at 304M, 32.96; avsolatorio/GIST-large-Embedding-v0 at 304M, 32.36; Omartificial-Intelligence-Space/Arabic-all-nli-triplet-Matryoshka at 86M, 32.35; sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 at 22M, 31.49; ibm-granite/granite-embedding-125m-english at 86M, 31.46; intfloat/e5-base-v2 at 86M, 31.44; Snowflake/snowflake-arctic-embed-s at 22M, 31.23; Omartificial-Intelligence-Space/Arabic-labse-Matryoshka at 86M, 31.02; thenlper/gte-base at 86M, 30.82; BAAI/bge-base-en-v1.5 at 86M, 30.78; jinaai/jina-embedding-s-en-v1 at 19M, 30.61; thenlper/gte-small at 22M, 30.07; Omartificial-Intelligence-Space/Arabic-MiniLM-L12-v2-all-nli-triplet at 22M, 29.50; avsolatorio/GIST-Embedding-v0 at 86M, 29.44; Snowflake/snowflake-arctic-embed-m at 86M, 29.16; intfloat/e5-small-v2 at 21M, 29.13; sdadas/mmlw-e5-large at 304M, 28.80; BAAI/bge-small-en-v1.5 at 22M, 28.73; dwzhu/e5-base-4k at 89M, 28.61; intfloat/e5-base at 86M, 28.61; intfloat/e5-small at 21M, 28.51; sentence-transformers/LaBSE at 86M, 28.49; abhinand/MedEmbed-small-v0.1 at 22M, 28.02; minishlab/potion-multilingual-128M at 0 (static), 27.99; Snowflake/snowflake-arctic-embed-m-v1.5 at 86M, 27.97; sentence-transformers/all-mpnet-base-v2 at 86M, 27.49; Gameselo/STS-multilingual-mpnet-base-v2 at 86M, 27.15; ibm-granite/granite-embedding-30m-english at 11M, 26.60; avsolatorio/GIST-small-Embedding-v0 at 22M, 26.37; Mihaiii/Ivysaur at 11M, 26.33; avsolatorio/GIST-all-MiniLM-L6-v2 at 11M, 26.20; sdadas/mmlw-roberta-large at 304M, 25.73; sentence-transformers/all-MiniLM-L12-v2 at 22M, 25.46; manu/sentence_croissant_alpha_v0.4 at 1.214B, 25.27; shibing624/text2vec-base-multilingual at 22M, 24.79; sentence-transformers/all-MiniLM-L6-v2 at 11M, 23.43; Snowflake/snowflake-arctic-embed-xs at 11M, 23.13; sdadas/mmlw-e5-base at 86M, 23.13; sergeyzh/LaBSE-ru-turbo at 87M, 22.97; sdadas/mmlw-roberta-base at 86M, 21.39; cointegrated/LaBSE-en-ru at 87M, 20.17; Mihaiii/Bulbasaur at 6M, 19.74; Mihaiii/Venusaur at 4M, 19.58; ai-forever/ru-en-RoSBERTa at 303M, 19.45; sdadas/mmlw-e5-small at 22M, 19.17; Mihaiii/Wartortle at 6M, 18.75; Mihaiii/gte-micro-v4 at 7M, 18.46; Mihaiii/Squirtle at 4M, 18.08; deepvk/USER-base at 388M, 16.41; consciousAI/cai-lunaris-text-embeddings at 304M, 13.54; sergeyzh/rubert-tiny-turbo at 3M, 7.51; Mihaiii/gte-micro at 6M, 7.13; silma-ai/silma-embeddding-matryoshka-v0.1 at 86M, 6.40; cointegrated/rubert-tiny at 3M, 5.35; cointegrated/rubert-tiny2 at 3M, 5.35; DeepPavlov/rubert-base-cased-sentence at 15M, 5.16; Omartificial-Intelligence-Space/Arabic-mpnet-base-all-nli-triplet at 86M, 4.79; Omartificial-Intelligence-Space/Arabert-all-nli-triplet-Matryoshka at 86M, 4.75; DeepPavlov/distilrubert-small-cased-conversational at 15M, 2.29; DeepPavlov/rubert-base-cased at 1.188B, 2.14; ai-forever/sbert_large_mt_nlu_ru at 304M, 2.08; Omartificial-Intelligence-Space/Marbert-all-nli-triplet-Matryoshka at 86M, 1.87; ai-forever/sbert_large_nlu_ru at 304M, 1.87; deepvk/deberta-v1-base at 85M, 1.09. On the Pareto frontier (nothing is both smaller and higher-scoring): mteb/baseline-bm25s, hotchpotch/bekko-embedding-v1-a8m, hotchpotch/bekko-embedding-v1-a25m, Snowflake/snowflake-arctic-embed-m-v2.0, Snowflake/snowflake-arctic-embed-l-v2.0, malteos/most-embed-de. Contributed by malteos. Source: MTEB, leaderboard.mteb.org.

031 3

Reposted by Kenneth Enevoldsen

MTEB @mteb.org · 25/08/2026

New benchmark: BRIGHT-Pro, a new multi-aspect retrieval benchmark on real StackExchange questions across 7 domains, built for agentic search — scored on surfacing several reasoning aspects, not one passage. AQ-MedAI/Diver-Retriever-4B-1020 leads at 55.88. Contributed by yilunzhao.

New benchmark: BRIGHT-Pro — A new multi-aspect retrieval benchmark on real StackExchange questions. Bar chart of the top 10 models on BRIGHT-Pro (7 tasks), out of 19 models with complete results. 1. AQ-MedAI/Diver-Retriever-4B-1020, 55.88, 3.6B parameters; 2. AQ-MedAI/Diver-Retriever-4B, 51.97, 3.6B parameters; 3. AQ-MedAI/Diver-Retriever-0.6B, 48.99, 440M parameters; 4. AQ-MedAI/Diver-Retriever-1.7B, 48.02, 1.4B parameters; 5. yale-nlp/RTriever-4B, 46.88, 3.6B parameters; 6. Qwen/Qwen3-Embedding-8B, 41.75, 6.9B parameters; 7. Qwen/Qwen3-Embedding-4B, 40.84, 3.6B parameters; 8. Alibaba-NLP/gte-Qwen2-7B-instruct, 37.75, 6.5B parameters; 9. ReasonIR/ReasonIR-8B, 32.94, 7.0B parameters; 10. Qwen/Qwen3-Embedding-0.6B, 29.93, 440M parameters. Contributed by yilunzhao. Source: MTEB, leaderboard.mteb.org.

121 2

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 24/08/2026

I actually not sure it is clear that it is illegal. It follows an opt-out model (which as I understand might be compliant with Europes text and data mining regulation). So far we haven't seen any decisions on it so it all speculations. I agree though, openly licensed content has many benefits

120 2

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 22/08/2026

Danish dynaword was ~4B tokens we released it. Less than a year after it is now more than doubled in size 🔥 Really shows the power of building externsible open datasets.

0111 11

Reposted by Kenneth Enevoldsen

MTEB @mteb.org · 18/08/2026

1/ MTEB's audio and video coverage grew substantially this month — 25 new tasks across 16 pull requests, spanning music, speech, environmental sound, and video retrieval. 10 of the 16 cite a published paper.

MTEB expands: 25 new audio & video tasks. 16 pull requests, 20–30 Jul 2026. 18 audio tasks, 7 video tasks, across 16 pull requests. 10 of 16 cite a published paper. Featured: Music — Song Describer: Text-to-music retrieval: the query is a caption and the goal is to retrieve the music track it describes from the corpus of 706 recordings. (Manco et al. (2023)) Speech — SPEECH-COCO: Queries are spoken captions and the corpus contains the images; the goal is to retrieve the image described by the spoken caption. (Havard et al. (2017)) Environmental sound — SoundingEarth: Queries are field recordings and the corpus contains the aerial images; the goal is to retrieve the image of the location where the recording was made. (Heidler et al. (2023)) Audio scene — CASTELLA: Retrieval on CASTELLA, the human-annotated long-audio benchmark used as the DCASE 2026 Task 6 evaluation set: given a caption describing a moment, retrieve the full recording (1 to 5 minutes) that contains it. The corpus is the 566 complete recordings and there are 1,347 caption queries, each with its source recording as the single relevant document. Adapted from CASTELLA's audio moment localization task, which regresses a continuous time interval, to whole-recording retrieval for the mteb ranking format. (Munakata et al. (2025)) Video retrieval — MomentSeeker: Queries combine a text question with a reference image, e.g. asking what happened around the pictured moment. (Yuan et al. (2025)) Video QA — MMVU: MMVU is an expert-level, multi-discipline video understanding benchmark with questions spanning 27 subjects across Science, Healthcare, Humanities & Social Sciences, and Engineering. Each multiple-choice example pairs a specialized-domain video with a question and 5 candidate answers. The task is formulated as multiple-choice retrieval: given the (video, question) pair, retrieve the correct candidate. Used the public validation multiple-choice subset (~625 examples). (Zhao et al. (2025)) Credit by task count: dukesun99 (14 tasks) · wissam-sib (9) · Rakshitha-Ireddi (2). Source: MTEB, leaderboard.mteb.org.

121 2

Reposted by Kenneth Enevoldsen

MTEB @mteb.org · 12/08/2026

New on MTEB: webAI-Official/webAI-ColVec1.1-8b, a visual document retrieval model, takes 1st of 41 on ViDoRe(v3) at 64.95. Its 4B sibling places 3rd at 63.90 — and both beat the lab's previous generation, which sits 6th and 7th.

New model on MTEB: webAI ColVec1.1 — 1st and 3rd on ViDoRe(v3) — both ahead of the lab's own prior generation. Bar chart of the top 10 models on ViDoRe(v3) (10 tasks), out of 41 models with complete results. 1. webAI-Official/webAI-ColVec1.1-8b, 64.95, 8.4B parameters (this release); 2. vultr/VultronRetrieverPrime-Qwen3.5-8B, 64.26, 8.4B parameters; 3. webAI-Official/webAI-ColVec1.1-4b, 63.90, 4.5B parameters (this release); 4. vultr/VultronRetrieverCore-Qwen3.5-4.5B, 63.57, 4.5B parameters; 5. nvidia/nemotron-colembed-vl-8b-v2, 63.42, 8.7B parameters; 6. webAI-Official/webAI-ColVec1-9b, 63.00, 9.4B parameters (same lab, previous generation); 7. webAI-Official/webAI-ColVec1-4b, 62.22, 4.5B parameters (same lab, previous generation); 8. TomoroAI/tomoro-colqwen3-embed-8b, 61.59, 8.0B parameters; 9. nvidia/nemotron-colembed-vl-4b-v2, 61.54, 4.8B parameters; 10. athrael-soju/colqwen3.5-4.5B-v3, 61.46, 4.6B parameters. Both ColVec1.1 sizes outrank the ColVec1 generation they replace. Contributed by zhanlunchang-webai. Source: MTEB, leaderboard.mteb.org.

262 6

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 10/08/2026

While we see many open weights model these days we are often still missing training code and training data, making it hard for the field to iterate and improve upon the works of others

070 7

Reposted by Kenneth Enevoldsen

MTEB @mteb.org · 07/08/2026

New on MTEB: Hanno-Labs/dinghy-law-4b-v1, a legal-domain embedding model, takes 1st on MTEB(Law, v1) at 71.22. Its 0.6B sibling places 4th at 65.83 — the smallest model in the top eight by an order of magnitude.

New model on MTEB: dinghy-law — 1st and 4th on MTEB(Law, v1). Bar chart of the top 8 models on MTEB(Law, v1) (8 tasks), out of 159 models with complete results. 1. Hanno-Labs/dinghy-law-4b-v1, 71.22, 4.0B parameters; 2. Mira190/Euler-Legal-Embedding-V1, 70.37, 7.6B parameters; 3. minetta/nemotron-3-embed-8b-legal, 69.33, 8.0B parameters; 4. Hanno-Labs/dinghy-law-0.6b-v1, 65.83, 0.6B parameters; 5. voyageai/voyage-law-2, 65.39, proprietary; 6. codefuse-ai/F2LLM-v2-14B, 64.66, 14B parameters; 7. voyageai/voyage-3, 64.13, proprietary; 8. infly/inf-retriever-v1, 63.68, 7.1B parameters. Contributed by trashhalo. Source: MTEB, leaderboard.mteb.org.

021 2

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 05/08/2026

Many use hard negatives from larger models — so I am not sure it is clear cut that it is wasted, but def. true that more effective models exists

000 0

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 05/08/2026

Embedding models keep getting cheaper - what required ~300M parameters three years ago now requires only 11M. In practice, that means that what runs on the server today could run in the browser tomorrow.

0222 22

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 03/08/2026

dyna=dynamic mean that it continually changes. Same as a dyno in bouldering it stands in contrast to static. The word is an iteration on gigaword and simply denote that it contains text/words.

010 1

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 31/07/2026

Looking forward to seeing how people will use it! If you have any criticism or stuff to add do let me know

000 0

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 31/07/2026

Fewer than 40% of open-weight models share training code or training data. This makes it impossible to improve the training script or iterate on the data, and reduces transparency. To combat this, we previously introduced an openness score. Today we add it to our leaderboard.

150 5

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 24/07/2026

MTEB reached 5000 contributions! Happy to see the package expand, as well as the community around it.

050 5

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 22/07/2026

Of course, a teaser for upcoming work, but if you want to learn more, you can find the Dynaword paper presented at LREC 2026 here: huggingface.co/papers/2508.... If you just want the datasets, they are also publicly available already; you can find them here: huggingface.co/collections/...

huggingface.co

Paper page - Dynaword: From One-shot to Continuously Developed Datasets

Join the discussion on this paper page

020 2

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 15/07/2026

Of course it shouldn’t be required to read in the review, the main paper should be understandable either code nor appendix - but I think not including code is as odd as stating that “we will include appendix upon paper acceptance”. Especially if we as reviewers has to rate the value of the software

000 0

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 15/07/2026

I don’t believe it is that assumed - only 46% of papers claim to open source their code [1], but even if they state they will share upon acceptance, shouldn’t the code be a part of the reviewed article? — less than half of repositories examined in [2] contains the required information to reproduce.

110 1

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 14/07/2026

I understand that this is what people imagine the dataset or software would be if it were shared, and meta-reviewers seem to simply accept it -- Am I missing something? To me it seems like we are asking for a replication crisis #EMNLP, #ACL

130 3

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 14/07/2026

A somewhat standard reviewing practice keeps frustrating me. When a paper does not mention sharing code or data it seems common that people rate if it was shared. Giving it scores like: "Software: 4 = Useful: I would recommend [...]" "Dataset: 2 = Documentary: The new datasets will be [...]"

110 1

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 07/07/2026

I would say I generally like the idea, but would really love to see an analysis on who it would affect. Also seems like you could game it by simply adding additional authors, which seems like a problematic incentive

010 1

Reposted by Kenneth Enevoldsen

Vladimir Salnikov @v4ldelund.bsky.social · 05/07/2026

Not so long ago we expanded Dynaword to Swedish, Norwegian and Icelandic Now have a new addition - Faroese 🇫🇴 If you are working on Faroese NLP or know someone who works on it, we would greatly appreciate any advice regarding what datasets we should add -> huggingface.co/datasets/dan...

huggingface.co

danish-foundation-models/faroese-dynaword · Datasets at Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

2132 13

Kenneth Enevoldsen @kennethenevoldsen.eurosky.social · 03/07/2026

Really cool work from our Danish foundation model, from the group at SDU. Do check it out!

0121 12