Records · kv:2610.00026v1

Reproduction of hiyouga/LlamaFactory: Qwen3-4B LoRA supervised fine-tuning example

K-Veritas Team

Submitted by K-Veritas Team

Published 2026-10-08

Independent reproduction · Artifact evaluation · Natural language processing

PaperReproduction of: original code

Abstract

Ran the shipped example examples/train_lora/qwen3_lora_sft.yaml at the pinned commit unchanged except for the local model path and output directory: LoRA (rank 8, all linear layers) supervised fine-tuning of Qwen3-4B-Instruct-2507 on the bundled identity and alpaca_en_demo datasets (up to 1,000 samples each), 3 epochs, batch 1 with 8 accumulation steps, learning rate 1e-4 cosine, bf16. torchaudio was replaced with the CUDA 12.8 build matching PyTorch. One NVIDIA RTX 5060 Ti.

Sealed report

Cite this record

@misc{kveritas261000026v1,
  title={Reproduction of hiyouga/LlamaFactory: Qwen3-4B LoRA supervised fine-tuning example},
  author={K-Veritas Team},
  year={2026},
  howpublished={K-Veritas Records},
  note={kv:2610.00026v1},
  url={https://kveritas.org/records/2610.00026v1},
}

References

1. [1] Mamadou K. Keita and Christopher Homan. Computer Science Conferences Should Require Nonrepudiable Experimental Results. NeurIPS 2026 Position Paper Track. arXiv:2605.08586.

Sealed with K-Veritas [1]. Listed, not endorsed. A record shows these results came from the sealed code, unchanged. Whether the work is sound is for the reader to judge.