OpenEuroLLM 9B math-RL canary, step 16 (experimental)

This is the 16-update starting checkpoint for the separate oellm-9b-math-rlvr-phase-a-32step-experimental release. It is a full BF16 model export, not an adapter. It starts from Neonkraft/oellm-9b-256k-theta64m-prelude-anneal300b-instruct-sft at revision 85bf18fb4f0bee6ac6270f06b1d1c6b3be200f31.

This is research evidence that a small, verifier-backed RL run can complete on LUMI. It is not an official OpenEuroLLM release, a general math benchmark winner, or evidence of safety or 256K-context quality. It was trained with TMAX/Open-Instruct DAPO, not Verl. The model.safetensors file is 18,203,942,400 bytes; SHA-256: c5b01a4b13d5bd76fdb4a69e73994e73704bb73e62a6a067068032f46a3f7565.

Training

• Source: birgermoell/oellm-math-rlvr at revision 0ffc9d6dc82717c25733b3172f4dbd63e48bab68.
• Training pool: 256 English difficulty-1-to-3 problems. The exact selected Parquet is training_data/english-pilot.parquet, SHA-256 ec33cc0fcbd6be73d87ed8f0582683cc6f2e4e56297ffa7db92f3712a5ec3aff.
• Algorithm: DAPO with outcome-verifiable math reward; 16 optimizer updates, eight unique prompts by eight stochastic samples per update, temperature 1.0, active sampling.
• Optimizer: constant learning rate 5e-7, beta=0, BF16, DeepSpeed ZeRO-3, gradient checkpointing; seed 42.
• Topology: two LUMI-G nodes, eight MI250X GCDs for the learner and eight for rollout engines. Job 22398668 completed in 14m30s, about 3.87 allocated GCD-hours.
• Exact control/config: BirgerMoell/oellm-rlvr and repro/canary-run.yaml (SHA-256 c55f686197dbbee6961a3f2f734b488a9379f566e21c54fb7116c001864f0974). Backend: OpenEuroLLM/tmax-reproduction at 3f80d37042402b8363f39c9535723b0d4cb8de54.

The YAML records LUMI-specific absolute paths and container image. To rerun, replace the project account, checkpoint and data paths for your allocation while preserving the recorded data hashes and training parameters. ground_truth and verifier metadata must never be included in model-visible prompts.

Evidence and limits

All 16 updates had finite, nonzero gradients; 128/128 accepted prompt groups had mixed rewards, and a cold restart from saved learner state completed an additional update. On a frozen 1,024-prompt English procedural holdout, greedy exact-answer accuracy was 42.97% for the SFT parent and 44.63% for this canary (+1.66 percentage points; paired bootstrap 95% interval +0.20 to +3.13 points). This holdout is drawn from the same procedural generator collection; step 16 was selected after inspecting step-8 and step-16 results, so the interval is exploratory rather than an untouched confirmation. This is not an independent general-math estimate. See the canary qualification record.

The model can still be wrong, overfit procedural templates, or produce unbalanced or unsound reasoning. Think-tag structure is not proof of correct reasoning. No safety, code, agentic, or long-context retention evaluation is claimed here.

License and attribution

Apache-2.0, following the parent model and the math-source dataset. The SFT parent was trained on Dolci-Instruct-SFT, whose separate data terms and attribution are described on its model card. OpenEuroLLM and LUMI/CSC contributors are acknowledged; this personal-account release does not imply project endorsement.