⚗️ Style Transfer: new hyperparameters and modify experiments

arkhn · Oct 16, 2024 · 373311f · 373311f
1 parent bbbc3b0
commit 373311f
Show file tree

Hide file tree

Showing 5 changed files with 17 additions and 4 deletions.
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
@@ -0,0 +1,12 @@
+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=1500 \
+            score.model.model_name_or_path=sentence-transformers/all-mpnet-base-v2 \
+            dataset.sft_ratio=0.06 \
+            dataset.gen_ratio=0.7 \
+            sft.training_args.eval_steps=30 \
+            score.train.train_size=0.3 \
+            dataset.sft_dataset.size=300
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
@@ -1,3 +1,4 @@
+export CUDA_VISIBLE_DEVICES=0
 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 \
@@ -6,5 +7,6 @@ python style_transfer/run_rb_gen.py model.name=meta-llama/Llama-3.2-3B-Instruct
             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
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 +1,2 @@
+export CUDA_VISIBLE_DEVICES=0
 python style_transfer/run_rb_gen.py
diff --git a/lib/style-transfer/configs/rb_gen/dpo/default.yaml b/lib/style-transfer/configs/rb_gen/dpo/default.yaml
@@ -5,11 +5,10 @@ training_args:
   save_steps: 50
   gradient_accumulation_steps: 16
   gradient_checkpointing: false
-  learning_rate: 2e-6
+  learning_rate: 5e-7
   weight_decay: 1e-7
   eval_strategy: "no"
-  num_train_epochs: 10
-  output_dir: "models/dpo"
+  num_train_epochs: 5
   optim: "adafactor"
   save_only_model: true
   remove_unused_columns: false

diff --git a/lib/style-transfer/configs/rb_gen/gen/default.yaml b/lib/style-transfer/configs/rb_gen/gen/default.yaml
@@ -6,7 +6,6 @@ sampling_params:
   top_p: 1.0
 llm:
   _target_: vllm.LLM
-  model: "models/merged/"
   speculative_model: "[ngram]"
   num_speculative_tokens: 5
   ngram_prompt_lookup_max: 4