OpenEuroLLM 9B post-training

How the OpenEuroLLM 9B 256K base model is post-trained into a multilingual assistant on the Leonardo supercomputer: Stage 1 Instruct SFT → Stage 2 Reasoning SFT → Stage 3 Tool-use SFT → Stage 4 DPO → Stage 5 RL. The documents below are the working record of the run. They are updated as it progresses, and are written so the pipeline can be reproduced.

Document	What it contains
Plan and decisions	goals, status dashboard, dated decisions with reasons, stage plans, evaluation protocol, budget, risks, next steps
Playbook	the reproducible record: data, configs, hashes, job IDs, costs, results, pitfalls, exact commands
RL run program	Stage 5 run by run: questions, run cards, results, stop rules, merging policy
Stage 5 data and rewards	RL data pools with licenses, rewards, difficulty windows, environment tiers

Code: qwen35-posttrain (SFT/DPO data builders, trainers, Slurm launchers, evaluation scripts) and oellm-rlvr (RL control plane). Related benchmark: EUROPA Reasoning.

Paths in the documents use placeholders ($WORK, $SCRATCH, <account>); model checkpoints are private until evaluation and license review.