MarinSkyRL Open-MOPD student, step 32
This is a MarinSkyRL-native Open-MOPD student after 32 optimizer steps. It starts from the authors' mixed-domain SFT checkpoint. Student responses were scored by the authors' math, code, and instruction-following RL teachers, routed by domain. The objective uses the student's selected top-16 token IDs and a clipped policy surrogate. This is an early checkpoint, not the authors' step-200 final model.
The checkpoint is an unquantized, six-file Hugging Face export of the durable MarinSkyRL global_step_32 FSDP2 checkpoint. The policy export was used for the independent step-32 evaluation. The export's model.safetensors SHA-256 is bb7326640142069bc2e1fba5f54f15e0cccb1ff861f34f318b372eaab7abaf4b. The original trainer-state SHA-256 is a3f50a849ed43f06deef731bc676a1bf8d31fcf5d34ed2f63c8b6aa1c985da34. The training source is MarinSkyRL commit 12e6da9e.
The retained training and inline-evaluation traces cover the original, fast-gate, and fast-full run segments. They include student responses and teacher-route metadata, but not the historical per-token teacher scoring arrays. The local open-mopd-repro/open-mopd-reproduce.md artifact gives the exact source, input, and launch sequence.
| Independent evaluation | Correct / outputs | Accuracy |
|---|---|---|
| AIME 2024, mean@64 | 423 / 1,920 | 22.03% |
| AIME 2025, mean@64 | 448 / 1,920 | 23.33% |
| IFEval, mean@1 | 403 / 541 | 74.49% |
These are three of the six released target benchmarks. LiveCodeBench v5/v6 and IFBench were generated, but paper-comparable scores are not claimed here. The training run reached a durable step-34 checkpoint and was then stopped by the campaign owner. This model is the earlier evaluated step-32 export. See the Open-MOPD paper for the authors' full experiment and final results.
Use AutoTokenizer.from_pretrained and AutoModelForCausalLM.from_pretrained with this repository ID. The export contains config.json, generation_config.json, tokenizer.json, tokenizer_config.json, and chat_template.jinja with the weights. For comparable AIME results, use the released 30-question sets with 64 samples per question and the recorded evaluation prompts, generation limits, and grader. A single greedy answer is not comparable to mean@64.