From 39984f73c292f7036533e1df95ee074ee05dddad Mon Sep 17 00:00:00 2001 From: Ethanfel Date: Mon, 6 Apr 2026 00:05:16 +0200 Subject: [PATCH] docs: add observed batching results to training guide Co-Authored-By: Claude Sonnet 4.6 --- LORA_TRAINING.md | 2 ++ 1 file changed, 2 insertions(+) diff --git a/LORA_TRAINING.md b/LORA_TRAINING.md index 3d32605..cac7fc8 100644 --- a/LORA_TRAINING.md +++ b/LORA_TRAINING.md @@ -199,6 +199,8 @@ The table below gives a rough scaling guide. Quality and diversity of recordings Higher batch size gives smoother loss curves and faster convergence. If you have headroom, prefer larger batches over more steps. +**Observed results:** batch 16 reaches the same loss in ~2600 steps that batch 1 needed 8000+ steps to reach, with a near-perfectly smooth curve. On a 24 GB GPU, batch 16 is the recommended default for `large_44k`. + ### Rank | Rank | Use case |