From 555023532fe86e7d2944047da27f358a5ce88648 Mon Sep 17 00:00:00 2001 From: Masahiko AMANO Date: Thu, 21 May 2026 21:18:37 +0300 Subject: [PATCH] scripts: update fine-tune defaults to lr=3e-5, epochs=30 Matches the configuration that produced finetuned.pt (val ppl 2.15, best epoch 20, early stopped at 30). Co-Authored-By: Claude Sonnet 4.6 --- scripts/train.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/scripts/train.py b/scripts/train.py index e267927..143d8a1 100644 --- a/scripts/train.py +++ b/scripts/train.py @@ -55,10 +55,10 @@ TRAIN_CFG = TrainConfig( output=CHECKPOINT, init_from=INIT_FROM, # Small corpus (~45 train files) → ~6 batches/epoch. - # 50 epochs × 6 = ~300 gradient steps; patience=10 gives a 60-step window. - epochs=50, + # 30 epochs × 6 = ~180 gradient steps; patience=10 gives a 60-step window. + epochs=30, batch_size=8, - lr=1e-5, + lr=3e-5, warmup_steps=10, patience=10, seed=42,