Huikang Weighted Interleaved 90/10 30B LoRA 20260614

• Hugging Face: https://huggingface.co/matthewagi/huikang-lb086-weighted-interleaved-90-10-30b-lora-20260614
• Kaggle dataset: https://www.kaggle.com/datasets/themattness/hk-wi9010-30b-lora-20260614

Contents

• adapter_config.json
• adapter_model.safetensors
• run_config.json
• final_eval_metrics.json
• split_summary.json
• train_launcher.sh
• train_log_tail.txt
• upload_manifest.json

Training Provenance

• Base model: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
• LoRA rank/alpha: r32/alpha32
• Train rows: 7756
• Eval rows: 859
• Effective global batch size: 32
• Source output dir: /workspace/outputs/huikang-lb086-weighted-interleaved-90-10-b300-30b-r32a32-mb16ga2-20260614_162726
• Adapter selection: final root adapter after step 243 training.

Final Eval