⚗️ Style Transfer: score hp changes

arkhn · Oct 18, 2024 · 754acb2 · 754acb2
1 parent eebebb0
commit 754acb2
Show file tree

Hide file tree

Showing 5 changed files with 17 additions and 14 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
@@ -1,4 +1,4 @@
-export CUDA_VISIBLE_DEVICES=1
+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 \
@@ -10,5 +10,6 @@ python style_transfer/run_rb_gen.py model.name=meta-llama/Llama-3.2-3B-Instruct
             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
+            dpo.training_args.num_train_epochs=40 \
+            dpo.percentile=70 \
+            score.batch_size=64
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
@@ -1,4 +1,4 @@
-export CUDA_VISIBLE_DEVICES=1
+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 \
@@ -10,5 +10,6 @@ python style_transfer/run_rb_gen.py model.name=meta-llama/Llama-3.2-3B-Instruct
             sft.training_args.eval_steps=30 \
             score.train.train_size=0.3 \
             dataset.sft_dataset.size=300 \
-            dpo.training_args.num_train_epochs=80 \
-            dpo.percentile=70
+            dpo.training_args.num_train_epochs=40 \
+            dpo.percentile=70 \
+            score.batch_size=64
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
@@ -1,4 +1,4 @@
-export CUDA_VISIBLE_DEVICES=0
+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 \
@@ -10,5 +10,6 @@ python style_transfer/run_rb_gen.py model.name=meta-llama/Llama-3.2-3B-Instruct
             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
+            dpo.training_args.num_train_epochs=40 \
+            dpo.percentile=70 \
+            score.batch_size=64
diff --git a/lib/style-transfer/bash/experiment/rb_gen/az/test-azure.sh b/lib/style-transfer/bash/experiment/rb_gen/az/test-azure.sh
@@ -1,2 +1,2 @@
-export CUDA_VISIBLE_DEVICES=0
+export CUDA_VISIBLE_DEVICES=$1
 python style_transfer/run_rb_gen.py
diff --git a/lib/style-transfer/configs/rb_gen/score/default.yaml b/lib/style-transfer/configs/rb_gen/score/default.yaml
@@ -6,10 +6,10 @@ model:
   model_name_or_path: "sentence-transformers/all-mpnet-base-v2"
 
 train:
-  warmup_steps: 50
-  use_ground_truth: false
-  epochs: 1
-  train_size: 0.5
+  warmup_steps: 10
+  use_ground_truth: true
+  epochs: 5
+  train_size: 0.3
   loss:
     _target_: sentence_transformers.losses.ContrastiveTensionLoss
     _partial_: true