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 <noreply@anthropic.com>
This commit is contained in:
+3
-3
@@ -55,10 +55,10 @@ TRAIN_CFG = TrainConfig(
|
|||||||
output=CHECKPOINT,
|
output=CHECKPOINT,
|
||||||
init_from=INIT_FROM,
|
init_from=INIT_FROM,
|
||||||
# Small corpus (~45 train files) → ~6 batches/epoch.
|
# Small corpus (~45 train files) → ~6 batches/epoch.
|
||||||
# 50 epochs × 6 = ~300 gradient steps; patience=10 gives a 60-step window.
|
# 30 epochs × 6 = ~180 gradient steps; patience=10 gives a 60-step window.
|
||||||
epochs=50,
|
epochs=30,
|
||||||
batch_size=8,
|
batch_size=8,
|
||||||
lr=1e-5,
|
lr=3e-5,
|
||||||
warmup_steps=10,
|
warmup_steps=10,
|
||||||
patience=10,
|
patience=10,
|
||||||
seed=42,
|
seed=42,
|
||||||
|
|||||||
Reference in New Issue
Block a user