OPENEUROLLM · RESULTS
Explore the datasets, models, training configurations, evaluation frameworks and research outputs produced by OpenEuroLLM — with direct links to the repositories and resources where they live.
Built in the open for researchers and developers.
Transparent · Multilingual · ReusableA structured catalogue of large-scale pre-training datasets with standardized metadata, language statistics and reproducible ingestion recipes for OpenEuroLLM and the wider EuroHPC ecosystem.
Project outputs
26 results
A structured catalogue of large-scale pre-training datasets with standardized metadata, language statistics and reproducible ingestion recipes for OpenEuroLLM and the wider EuroHPC ecosystem.
Open multilingual synthetic pre-training data created by translating high-quality English material into OpenEuroLLM target languages.
Aligned translated data released for selected languages as part of the MultiSynt synthetic-data effort.
A second aligned translated dataset published for selected languages as part of OpenEuroLLM's synthetic and multilingual data work.
A robust pipeline for massively multilingual parallel data derived from OPUS, covering cleaning, normalization, filtering, parallelism checks and quality evaluation, optimized for SLURM-based HPC.
Tooling and masking strategies used to process sensitive information in the first OpenEuroLLM training-data cycle, regex-based, adapted for HPC environments and multilingual scenarios.
LLM-based PII detection tooling adapted for HPC processing and used for the project's second training-data cycle.
A framework that records the exact dataset selection and preprocessing behind each model-training cycle, making the training-data recipe inspectable and reproducible.
A model that annotates text documents across many different categories and languages. Used to filter, select, and curate OpenEuro LLM training data.
Experimental reference models shared in collaboration with the Open-Sci project and used in OpenEuroLLM evaluation and scaling experiments.
A public collection of monolingual reference models shared in collaboration with HPLT and used as part of OpenEuroLLM's comparative work.
The public OpenEuroLLM model collection, including experimental model and data-mixture releases shared during development.
OpenEuroLLM's public Prelude model release and intermediate checkpoints, used as a major pre-production run for continuous evaluation and scaling-law validation.
The public OpenEuroLLM tokenizer released as part of the model and data preparation stack.
An open-source framework for large-scale experimentation on EuroHPC systems that generates SLURM scripts, supports nested hyperparameter sweeps and helps monitor training runs.
A public-facing space exposing live model-training progress so external communities can follow development as it happens.
Public Weights & Biases workspace used to share training logs and live metrics for the project's Prelude (aka baby) 9B dense training work.
Essential software identified by the project as part of the production stack behind MultiSynt's synthetic-data rollout.
A unified framework for reproducible LLM evaluations on EuroHPC clusters, with cluster-aware execution, reproducibility metadata, large job arrays and benchmark suites.
A standalone collection of OpenEuroLLM evaluation tasks that can be consumed by oellm-eval, the Evaluation Harness and other toolkits.
An evaluation framework focused on task variants, metrics and multilingual evaluation fidelity, complementing oellm-eval's orchestration layer.
A unified framework for running multiple LLM-judge benchmarks, including work on tuned open-weight judges, Soft-Elo and calibrated uncertainty.
A public benchmark collection supporting the project's effort to detect and prevent evaluation data from leaking into pre-training data.
Public tooling and resources supporting OpenEuroLLM's pre-training decontamination process.
The publication associated with the MultiSynt multilingual synthetic-data work, referenced in the project's technology-rollout section.
The project's public publications index. The first periodic report records 58 publications acknowledging OpenEuroLLM during the reporting period.
Built openly
OpenEuroLLM is publishing project outputs as they are built. Follow the links above to use the datasets, inspect the code, evaluate the models, reproduce experiments or contribute improvements upstream.