Joakimpalm-Zen has updated Xyntetik Runner to train LoRA adapters directly using the same quantized GGUF weights it serves for inference, eliminating the need for a separate FP16 training copy or runtime.

  • A tool-use LoRA was trained against a frozen Qwen3-4B Q4_K_M GGUF, achieving 1.00 accuracy on exact tool calls and right-tool selection in held-out evaluations.
  • Two independent runs with the same base GGUF, dataset, seed, and config produced byte-identical adapter GGUFs with matching SHA256 hashes.
  • Comparisons across BF16, Q8_0, and Q4_K_M quantizations showed that while all reached 1.00 accuracy, training through Q4 followed a measurably different optimization path (cosine similarity 0.9926) than the higher-precision versions.

The author provides full reproducibility artifacts on Hugging Face but notes the experiment is limited to a narrow synthetic task with small data, serving primarily as a demonstration of the training mechanism and deterministic properties rather than general capability improvement.