Safety and alignment evaluation
@BirgerMoell ci sta già lavorando.
Dal 12/10/2026.
Valutazione
Questa issue non è ancora stata valutata.
Descrizione
Multilingual safety and alignment evaluation of post-trained checkpoints
Goal
Evaluate whether SFT, DPO and subsequent post-training stages improve safety and alignment across OpenEuroLLM’s target languages, while identifying regressions that an overall score could hide.
Description
Use EUROPA Alignment to compare pinned checkpoints under the same generation and judging protocol.
Its public v0.1 suite covers 16 scenarios in each of 36 languages across refusal calibration, human wellbeing, privacy and security, high-stakes decisions, dangerous-capability boundaries, and agentic control.
Review individual answers, critical gate failures, over-refusal, malformed responses and the weakest languages alongside aggregate scores. The current DPO run provides initial evidence and a comparison with the SFT checkpoint.
The suite is a development pilot: 35 translated language packs await native review, and judge scores need human calibration. Report these limits with every comparison; do not treat the public suite as a release certification.
Deliverable scope
- Run the complete suite on the selected base/SFT, DPO and later release-candidate checkpoints using pinned model revisions and a consistent protocol.
- Publish reproducible results with per-language and per-domain comparisons, critical failures, over-refusal and representative answer-level evidence.
- Investigate material regressions and weak-language results; record whether they arise from model behavior, translation quality, output formatting or judge errors.
- Obtain native-speaker review and judge calibration for results used to support release decisions, and document remaining evaluation limits.
Dependencies
Exported, loadable checkpoints and evaluation compute; EUROPA Alignment and a pinned independent judge. Native review of translated scenarios and human calibration of judge decisions are required before making release-level multilingual safety claims.
- Lingua principale
- Nessun dato sulla lingua
- Stelle
- 3
- Fork
- 0
- Metriche di merge delle PR
- Nessuna PR unita negli ultimi 30g
Preparare l'ambiente
Questo progetto non fornisce container di sviluppo, Dockerfile né guida per i contributori, quindi l'ambiente è a tuo carico: parti dal suo README e consulta la nostra guida al primo contributo per i passaggi generali.
Come iniziare
- Leggi tutta la issue e poi la guida ai contributi del progetto.
- Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
- Fai un fork del repository e lavora su un branch.
- Apri una pull request che faccia riferimento al numero della issue.
Altre issue di OpenEuroLLM/Taskboard
-
[nucleus sparse] Experiment setupForse già presa @aaronkl l’ha presa 9 giorni fa. ApertaT4.1 - Optimization HPC T4.2 - model exploration
OpenEuroLLM/Taskboard#439 · 1 commento · 1 reazione · 1 assegnatario ·
-
Qwen3.5 35B-A3B hybrid MoE weak and strong scalingForse già presa @ibaucells l’ha presa 9 giorni fa. ApertaT4.1 - Optimization HPC
OpenEuroLLM/Taskboard#438 · 2 commenti · 1 assegnatario ·
-
Synthetic Generation for Multilingual Instruction TuningForse già presa @abhash-er l’ha presa 10 giorni fa. Aperta4.6 post-training
OpenEuroLLM/Taskboard#437 · 1 assegnatario ·
-
Finding optimal dataset mixtures for DPOForse già presa @geoalgo l’ha presa 11 giorni fa. Aperta4.6 post-training
OpenEuroLLM/Taskboard#435 · 3 assegnatari ·
-
DPO checkpoint mergingForse già presa @HannanJaved l’ha presa 11 giorni fa. Aperta4.6 post-training
OpenEuroLLM/Taskboard#434 · 2 commenti · 1 assegnatario ·