GRPO checkpoints — smu_20260905
Private archive of model exports at optimizer steps 10, 20, 30, 40 and 50 from
the smu_20260905 full201-std experiment. Each step_N folder contains the
complete BF16 model export, tokenizer, configuration and saved chat template.
The saved model files are retained unchanged. Training optimizer and sampler
states remain in the separate full training checkpoints on the source server.
Warm start: aaravshar/qwen36-35b-su-t3-practice-step160,
revision a1227bed431e535fb4d4dd9fa0d35bed98220587 (SFT step 160).
Architecture in the exported config: Qwen3_5MoeForConditionalGeneration.
The experiment uses the Qwen3.6-35B-A3B label and a harness-specific renderer.
Download one checkpoint:
from huggingface_hub import snapshot_download
snapshot = snapshot_download(
repo_id="aaravshar/qwen36-35b-su-t3-grpo-smu-20260905",
allow_patterns=["step_10/*"],
token=True,
)
checkpoint_directory = snapshot + "/step_10"
Replace step_10 with step_20, step_30, step_40 or step_50 as needed.
Use checkpoint_directory wherever the existing inference setup expects the
exported model path. The archive does not alter the saved prompt formatting.
export_manifest.json identifies the folders and export sizes; SHA256SUMS
records verified file hashes once the upload has completed.