Grug 67B A2B GLM53 RLVR1 Async · step 48
This is the Hugging Face safetensors export of optimizer step 48 from the active GLM53 RLVR1 asynchronous Snowball run. It starts from Grug-67B-A2B-GLM53-RLVR-SFT-2026.09.23 and was trained with MarinSkyRL and Megatron using router replay.
Checkpoint selection
Among the evaluated, saved checkpoints available when the export was selected (through optimizer step 52), steps 48 and 50 tied for the highest configured holdout eval/all/pass_at_1 at 0.53535. Step 48 had the higher holdout eval/all/avg_score: 0.58757 versus 0.58745 at step 50. These are run-internal holdout results, not independent benchmark claims. The training run was still active when this snapshot was chosen.
Provenance
- Training run:
/benfeuer/glm53-rlvr1-marin-coord-v5/users-benfeuer-checkpoints-glm53-rlvr1-async-2026.09.23-5509d1e0cef2 - Optimizer step: 48
- Source checkpoint:
s3://marin-us-east-02a/tmp/ttl=14d/skyrl/marin/users/benfeuer/checkpoints/glm53-rlvr1-async/2026.09.23/checkpoints/global_step_48 - Durable HF export:
s3://marin-us-east-02a/marin/users/benfeuer/checkpoints/glm53-rlvr1-async/2026.09.23/exports/global_step_48/policy - MarinSkyRL runtime:
a9e3066
Use the tokenizer and chat template shipped with this repository. The inherited OpenMDW 1.1 license applies.
The files use the Hugging Face safetensors layout and the custom grug_moe architecture. Inference requires a runtime with Grug support, such as Marin's vLLM fork; stock Transformers 5.12.1 does not recognize grug_moe without registering the model implementation.