⚗️ Style Transfer: modify hyperparameters to fix dpo learning problem

arkhn · Oct 17, 2024 · 0ca86de · 0ca86de
1 parent d3fa9b2
commit 0ca86de
Show file tree

Hide file tree

Showing 4 changed files with 22 additions and 3 deletions.
diff --git a/lib/style-transfer/bash/experiment/rb_gen/az/llama3.2-3b-complete.sh b/lib/style-transfer/bash/experiment/rb_gen/az/llama3.2-3b-complete.sh
@@ -0,0 +1,14 @@
+export CUDA_VISIBLE_DEVICES=1
+python style_transfer/run_rb_gen.py model.name=meta-llama/Llama-3.2-3B-Instruct \
+            model.peft_config.target_modules='["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"]' \
+            dataset.name=bio-datasets/mimic_style_transfer \
+            max_steps=5 \
+            dataset.num_generated_samples=3500 \
+            score.model.model_name_or_path=sentence-transformers/all-mpnet-base-v2 \
+            dataset.sft_ratio=0.06 \
+            dataset.gen_ratio=0.7 \
+            dataset.sft_dataset=null \
+            sft.training_args.eval_steps=30 \
+            score.train.train_size=0.3 \
+            dpo.training_args.num_train_epochs=80 \
+            dpo.percentile=70
diff --git a/lib/style-transfer/bash/experiment/rb_gen/az/llama3.2-3b-pb-seed.sh b/lib/style-transfer/bash/experiment/rb_gen/az/llama3.2-3b-pb-seed.sh
@@ -9,4 +9,6 @@ python style_transfer/run_rb_gen.py model.name=meta-llama/Llama-3.2-3B-Instruct
             dataset.gen_ratio=0.7 \
             sft.training_args.eval_steps=30 \
             score.train.train_size=0.3 \
-            dataset.sft_dataset.size=300
+            dataset.sft_dataset.size=300 \
+            dpo.training_args.num_train_epochs=80 \
+            dpo.percentile=70
diff --git a/lib/style-transfer/bash/experiment/rb_gen/az/llama3.2-3b.sh b/lib/style-transfer/bash/experiment/rb_gen/az/llama3.2-3b.sh
@@ -9,4 +9,6 @@ python style_transfer/run_rb_gen.py model.name=meta-llama/Llama-3.2-3B-Instruct
             dataset.gen_ratio=0.7 \
             dataset.sft_dataset=null \
             sft.training_args.eval_steps=30 \
-            score.train.train_size=0.3
+            score.train.train_size=0.3 \
+            dpo.training_args.num_train_epochs=80 \
+            dpo.percentile=70
diff --git a/lib/style-transfer/configs/rb_gen/dpo/default.yaml b/lib/style-transfer/configs/rb_gen/dpo/default.yaml
@@ -5,10 +5,11 @@ training_args:
   save_steps: 50
   gradient_accumulation_steps: 16
   gradient_checkpointing: false
-  learning_rate: 5e-7
+  learning_rate: 4e-6
   weight_decay: 1e-7
   eval_strategy: "no"
   num_train_epochs: 5
+  output_dir: "models/dpo/"
   optim: "adafactor"
   save_only_model: true
   remove_unused_columns: false