Fix hyperparameters in KTO example (#1474)

* Fix hparams in KTO example * Clean * Fix
huggingface · Mar 24, 2024 · 9ce7ac6 · 9ce7ac6
1 parent 99553c1
commit 9ce7ac6
Showing 1 changed file with 4 additions and 5 deletions.
diff --git a/examples/scripts/kto.py b/examples/scripts/kto.py
@@ -19,9 +19,9 @@
 # regular:
 python examples/scripts/kto.py \
     --model_name_or_path=gpt2 \
-    --per_device_train_batch_size 4 \
+    --per_device_train_batch_size 16 \
     --max_steps 1000 \
-    --learning_rate 1e-3 \
+    --learning_rate 2e-5 \
     --gradient_accumulation_steps 1 \
     --logging_steps 10 \
     --eval_steps 500 \
@@ -35,14 +35,13 @@
 # peft:
 python examples/scripts/kto.py \
     --model_name_or_path=gpt2 \
-    --per_device_train_batch_size 4 \
+    --per_device_train_batch_size 16 \
     --max_steps 1000 \
-    --learning_rate 1e-3 \
+    --learning_rate 2e-4 \
     --gradient_accumulation_steps 1 \
     --logging_steps 10 \
     --eval_steps 500 \
     --output_dir="kto_anthropic_hh" \
-    --optim rmsprop \
     --warmup_steps 150 \
     --report_to wandb \
     --bf16 \