Community/training - Revision history

Oe: /* Schedule */

2026-02-04T15:30:06Z

‎Schedule

Oe: /* Schedule */

2026-02-04T13:10:50Z

‎Schedule

Oe: /* Schedule */

2026-02-04T11:02:44Z

‎Schedule

Oe: /* Schedule */

2026-02-04T10:46:21Z

‎Schedule

Oe: /* Schedule */

2026-02-04T08:23:38Z

‎Schedule

Oe: /* Schedule */

2026-02-04T08:19:34Z

‎Schedule

Oe: /* Schedule */

2026-02-02T20:14:53Z

‎Schedule

Oe: /* Schedule */

2026-02-02T19:57:16Z

‎Schedule

Oe: /* Schedule */

2026-02-02T17:40:54Z

‎Schedule

Oe: /* Schedule */

2026-02-02T17:40:09Z

‎Schedule

@@ Line 99: / Line 99: @@
 |-
 | 21:00 || || '''Evening Session''':<br/>
-Javier de la Rosa, Rolv-Arild Braaten, Marthe Midtgaard, Angelina Zanardi: National Library of Norway<br/>
+[http://svn.nlpl.eu/outreach/skeikampen/2026/nb.pdf Javier de la Rosa, Rolv-Arild Braaten, Marthe Midtgaard, Angelina Zanardi: National Library of Norway]<br/>
 [http://svn.nlpl.eu/outreach/skeikampen/2026/openeurollm.pdf Sampo Pyysalo, Max Idahl, David Salias, Stephan Oepen, Shenbin Qian: OpenEuroLLM, MultiSynt]
 |}

@@ Line 114: / Line 114: @@
 | 10:00 || 10:30 || Coffee Break
 |-
-| 10:30 || 12:00 || '''Session 9''' Julia Kreutzer <p  class="mw-collapsible mw-collapsed">'''Optimizing data for multilingual post-training'''<br>In this session we will look into techniques for augmenting data collections for better multilingual coverage. We will discuss the role of translation and inference settings, and explore methods for optimizing multilingual data both on the prompt and the generation side.</p>
+| 10:30 || 12:00 || '''Session 9''' [http://svn.nlpl.eu/outreach/skeikampen/2026/kreutzer2.pdf Julia Kreutzer: Optimizing Data for Multilingual Post-Training] <p  class="mw-collapsible mw-collapsed">In this session we will look into techniques for augmenting data collections for better multilingual coverage. We will discuss the role of translation and inference settings, and explore methods for optimizing multilingual data both on the prompt and the generation side.</p>
 |-
 | 12:30 || 13:30 || Lunch

@@ Line 110: / Line 110: @@
 |colspan=3| Breakfast is available from 07:30
 |-
-| 08:30 || 10:00 || '''Session 8''' Barbara Plank <p  class="mw-collapsible mw-collapsed">'''NLP Beyond the Standard: Dialects, Variation, and Shared Representations in Multilingual Language Models'''<br>Multilingual language models have primarily focused on cross-lingual differences, with intra-language variation only recently gaining more attention. Dialects and non-standard varieties challenge core assumptions about data, representation, and evaluation. In this talk, I discuss what makes dialects particularly challenging for multilingual models, review approaches starting from early encoder-based methods, and give an overview of resources developed for dialectal NLP, with a focus on German dialects. I then turn to recent work on multilingual training dynamics and shared representations, analyzing when linguistic information and shared concept spaces emerge during training and where alignment breaks down. Although dialects are not yet explicitly modeled in this analysis, the findings provide insight into multilingual representation learning during pre-training. </p>
+| 08:30 || 10:00 || '''Session 8''' [http://svn.nlpl.eu/outreach/skeikampen/2026/plank.pdf Barbara Plank: NLP Beyond the Standard]<p  class="mw-collapsible mw-collapsed">'''Dialects, Variation, and Shared Representations in Multilingual Language Models'''<br>Multilingual language models have primarily focused on cross-lingual differences, with intra-language variation only recently gaining more attention. Dialects and non-standard varieties challenge core assumptions about data, representation, and evaluation. In this talk, I discuss what makes dialects particularly challenging for multilingual models, review approaches starting from early encoder-based methods, and give an overview of resources developed for dialectal NLP, with a focus on German dialects. I then turn to recent work on multilingual training dynamics and shared representations, analyzing when linguistic information and shared concept spaces emerge during training and where alignment breaks down. Although dialects are not yet explicitly modeled in this analysis, the findings provide insight into multilingual representation learning during pre-training. </p>
 |-
 | 10:00 || 10:30 || Coffee Break

@@ Line 90: / Line 90: @@
 |colspan=3| Free time (Lunch is available between 13:00 and 14:30)
 |-
-| 15:30 || 17:00 || '''Session 5''' [http://svn.nlpl.eu/outreach/skeikampen/2026/idahl.pdf Max Idahl: Multilingual Model-Based Quality Filtering for LLM Pretraining]<p  class="mw-collapsible mw-collapsed"><br>Data quality is the highest-leverage factor for LLM performance, with recent work showing significant training efficiency gains through careful curation. This presentation traces the evolution from rule-based filtering to modern model-based approaches that now work across dozens of languages. We cover the progression from basic perplexity-based filters, to FastText and encoder-based scorers, to our newly released Propella models that annotate documents across 18 properties for 57 languages at scale. The talk includes practical insights into building multilingual filtering pipelines.</p>
+| 15:30 || 17:00 || '''Session 5''' [http://svn.nlpl.eu/outreach/skeikampen/2026/idahl.pdf Max Idahl: Multilingual Model-Based Quality Filtering for LLM Pretraining]<p  class="mw-collapsible mw-collapsed">Data quality is the highest-leverage factor for LLM performance, with recent work showing significant training efficiency gains through careful curation. This presentation traces the evolution from rule-based filtering to modern model-based approaches that now work across dozens of languages. We cover the progression from basic perplexity-based filters, to FastText and encoder-based scorers, to our newly released Propella models that annotate documents across 18 properties for 57 languages at scale. The talk includes practical insights into building multilingual filtering pipelines.</p>
 |-
 | 17:00 || 17:20 || Coffee Break
 |-
-| 17:20 || 19:20 || '''Session 6''' David Salinas <p  class="mw-collapsible mw-collapsed">'''Challenges in Evaluating Generative Models'''<br>In this talk, we will discuss the evaluation of generative models, in particular Large Language Models (LLMs). Given that such models produce open-ended output, their evaluation requires different techniques than static evaluations such as simple question-answering benchmarks. We will first discuss human annotations and their use in leaderboards such as LMArena and ComparIA. We will then focus on automatic evaluation relying on LLM judges. In particular, we will describe current challenges with LLM judges before discussing their application in multilingual settings.</p>
+| 17:20 || 19:20 || '''Session 6''' [http://svn.nlpl.eu/outreach/skeikampen/2026/salinas.pdf David Salinas: Challenges in Evaluating Generative Models]<p  class="mw-collapsible mw-collapsed">In this talk, we will discuss the evaluation of generative models, in particular Large Language Models (LLMs). Given that such models produce open-ended output, their evaluation requires different techniques than static evaluations such as simple question-answering benchmarks. We will first discuss human annotations and their use in leaderboards such as LMArena and ComparIA. We will then focus on automatic evaluation relying on LLM judges. In particular, we will describe current challenges with LLM judges before discussing their application in multilingual settings.</p>
 |-
 | 19:30 ||  || Dinner

@@ Line 98: / Line 98: @@
 | 19:30 ||  || Dinner
 |-
-| 21:00 || || '''Evening Session''': National Library of Norway, OpenEuroLLM, MultiSynt
+| 21:00 || || '''Evening Session''':<br/>
 |}

@@ Line 86: / Line 86: @@
 |colspan=3 | Breakfast is available from 07:30
 |-
-| 09:00 || 10:30 || '''Session 4''' François Yvon <p  class="mw-collapsible mw-collapsed">'''Text Generation: Know your Options!'''<br>Text generation, contextual or non-contextual, is ubiquitous in the current LLM era, as it serves as the most basic block in multiple application contexts, from question answering and dialog systems to text summarization and machine translation, and many more. Generation is thus equally useful to compute deterministic and highly non-deterministic mappings with various level of output constraints. Furthermore, text generation is also used as a sub-routine of more complex generation strategies, aiming to produce syntactically well-formed (e.g. for code generation) or semantically consistent outputs, possibility through multiple steps of generation (e.g, in chain-of-thoughts generation) or to collect diverse samples from the generating distribution. To cover this considerable diversity of uses, multiple text generation strategies have been proposed, some less well-known than others. In this talk I will review various families of generation algorithms, from the most basic ones to the more sophisticated approaches, so as to document, as much as possible, the possible options that are available to text generation users. The final part will survey some decoding issues that are specific to multilingual models. </p>
+| 09:00 || 10:30 || '''Session 4''' [http://svn.nlpl.eu/outreach/skeikampen/2026/yvon2.pdf François Yvon: Text Generation: Know your Options!] <p  class="mw-collapsible mw-collapsed">Text generation, contextual or non-contextual, is ubiquitous in the current LLM era, as it serves as the most basic block in multiple application contexts, from question answering and dialog systems to text summarization and machine translation, and many more. Generation is thus equally useful to compute deterministic and highly non-deterministic mappings with various level of output constraints. Furthermore, text generation is also used as a sub-routine of more complex generation strategies, aiming to produce syntactically well-formed (e.g. for code generation) or semantically consistent outputs, possibility through multiple steps of generation (e.g, in chain-of-thoughts generation) or to collect diverse samples from the generating distribution. To cover this considerable diversity of uses, multiple text generation strategies have been proposed, some less well-known than others. In this talk I will review various families of generation algorithms, from the most basic ones to the more sophisticated approaches, so as to document, as much as possible, the possible options that are available to text generation users. The final part will survey some decoding issues that are specific to multilingual models. </p>
 |-
 |colspan=3| Free time (Lunch is available between 13:00 and 14:30)
 |-
-| 15:30 || 17:00 || '''Session 5''' Max Idahl <p  class="mw-collapsible mw-collapsed">'''Multilingual Model-Based Quality Filtering for LLM Pretraining'''<br>Data quality is the highest-leverage factor for LLM performance, with recent work showing significant training efficiency gains through careful curation. This presentation traces the evolution from rule-based filtering to modern model-based approaches that now work across dozens of languages. We cover the progression from basic perplexity-based filters, to FastText and encoder-based scorers, to our newly released Propella models that annotate documents across 18 properties for 57 languages at scale. The talk includes practical insights into building multilingual filtering pipelines.</p>
+| 15:30 || 17:00 || '''Session 5''' [http://svn.nlpl.eu/outreach/skeikampen/2026/idahl.pdf Max Idahl: Multilingual Model-Based Quality Filtering for LLM Pretraining]<p  class="mw-collapsible mw-collapsed"><br>Data quality is the highest-leverage factor for LLM performance, with recent work showing significant training efficiency gains through careful curation. This presentation traces the evolution from rule-based filtering to modern model-based approaches that now work across dozens of languages. We cover the progression from basic perplexity-based filters, to FastText and encoder-based scorers, to our newly released Propella models that annotate documents across 18 properties for 57 languages at scale. The talk includes practical insights into building multilingual filtering pipelines.</p>
 |-
 | 17:00 || 17:20 || Coffee Break

@@ Line 75: / Line 75: @@
 | 17:30 || 17:50 || Coffee Break
 |-
-| 17:50 || 19:20 || '''Session 3''' Julia Kreutzer <p  class="mw-collapsible mw-collapsed">'''Evaluating Generations Multilingually: Current challenges and Lessons from Machine Translation'''<br>In this session we will dive into the particular challenge of evaluating LLMs across many languages in generative tasks. We will take a look at the "sister field" of machine translation and inspect what principles have led to advances in understanding quality across languages. </p>
+| 17:50 || 19:20 || '''Session 3''' [http://svn.nlpl.eu/outreach/skeikampen/2026/kreutzer1.pdf Julia Kreutzer: Evaluating Generations Multilingually] <p  class="mw-collapsible mw-collapsed">'''Current Challenges and Lessons from Machine Translation'''<br>In this session we will dive into the particular challenge of evaluating LLMs across many languages in generative tasks. We will take a look at the "sister field" of machine translation and inspect what principles have led to advances in understanding quality across languages. </p>
 |-
 | 19:30 ||  || Dinner

@@ Line 67: / Line 67: @@
 | 13:00 || 14:00 || Lunch
 |-
-| 14:00 || 15:30 || '''Session 1''' Laurie Burchell and Pedro Ortiz Suarez <p  class="mw-collapsible mw-collapsed">'''Multilinguality at Common Crawl: improving language coverage for the largest open web corpus'''<br>The Common Crawl Foundation (CCF) provides the largest open corpus of web data, enabling a wide range of scientific and technical applications including large language model (LLM) development. However, our current data processing pipeline faces challenges when processing multilingual data, decreasing language representation and impacting downstream model performance. In this talk, we will discuss CCF’s initiatives to improve multilingual coverage and language identification of our web corpus. These efforts include soliciting crowd-sourced web seeds for under-served languages, running the First Workshop for Multilingual Data Quality Signals at COLM 2025, and creating CommonLID, a community-driven, human-annotated language identification benchmark for the web domain. Throughout, we emphasise the collaborative nature of our efforts, working in partnership with members of the NLP community to improve content available in their languages.</p>
+| 14:00 || 15:30 || '''Session 1''' [http://svn.nlpl.eu/outreach/skeikampen/2026/ortiz-burchell.pdf Laurie Burchell and Pedro Ortiz Suarez: Multilinguality at Common Crawl]<p  class="mw-collapsible mw-collapsed">'''Improving Language Coverage for the Largest Open Web Corpus'''<br>The Common Crawl Foundation (CCF) provides the largest open corpus of web data, enabling a wide range of scientific and technical applications including large language model (LLM) development. However, our current data processing pipeline faces challenges when processing multilingual data, decreasing language representation and impacting downstream model performance. In this talk, we will discuss CCF’s initiatives to improve multilingual coverage and language identification of our web corpus. These efforts include soliciting crowd-sourced web seeds for under-served languages, running the First Workshop for Multilingual Data Quality Signals at COLM 2025, and creating CommonLID, a community-driven, human-annotated language identification benchmark for the web domain. Throughout, we emphasise the collaborative nature of our efforts, working in partnership with members of the NLP community to improve content available in their languages.</p>
 |-
 | 15:30 || 15:50 || Coffee Break

@@ Line 71: / Line 71: @@
 | 15:30 || 15:50 || Coffee Break
 |-
-| 16:00 || 17:30 || '''Session 2''' [http://svn.nlpl.eu/outreach/skeikampen/2026/yvon1.pdf François Yvon: Evaluating Large LMs and their Multilingualism]<p  class="mw-collapsible mw-collapsed">'''Evaluating Multilingual Models'''<br> Large Language Models introduced in the recent years have been found extremely helpful to advance the state-of-the-art in many Natural Language Applications, notably due to their ability to compute numerical, high-dimensional, representations of linguistic units such as words or sentences. Multilingual language models go one step further and add the ability to handle multiple languages, sometimes even multiple scripts, with just one single model. In this presentation, I will discuss multilingual language models at length, with a focus on the evaluation of their multilingual abilities, which raises two difficult questions: (a) to evaluate their performance as if they were just a collection of monolingual models; (b) to evaluate their performance as integrated multilingual models, capable of bridging between languages. </p>
+| 16:00 || 17:30 || '''Session 2''' [http://svn.nlpl.eu/outreach/skeikampen/2026/yvon1.pdf François Yvon: Evaluating Large LMs and their Multilingualism]<p  class="mw-collapsible mw-collapsed">Large Language Models introduced in the recent years have been found extremely helpful to advance the state-of-the-art in many Natural Language Applications, notably due to their ability to compute numerical, high-dimensional, representations of linguistic units such as words or sentences. Multilingual language models go one step further and add the ability to handle multiple languages, sometimes even multiple scripts, with just one single model. In this presentation, I will discuss multilingual language models at length, with a focus on the evaluation of their multilingual abilities, which raises two difficult questions: (a) to evaluate their performance as if they were just a collection of monolingual models; (b) to evaluate their performance as integrated multilingual models, capable of bridging between languages. </p>
 |-
 | 17:30 || 17:50 || Coffee Break