From 754acb24eb49aa770d91b68ef53dc4b7d63f5e41 Mon Sep 17 00:00:00 2001 From: Simon Meoni Date: Fri, 18 Oct 2024 11:18:35 +0200 Subject: [PATCH] =?UTF-8?q?=E2=9A=97=EF=B8=8F=20Style=20Transfer:=20score?= =?UTF-8?q?=20hp=20changes?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../bash/experiment/rb_gen/az/llama3.2-3b-complete.sh | 7 ++++--- .../bash/experiment/rb_gen/az/llama3.2-3b-pb-seed.sh | 7 ++++--- .../bash/experiment/rb_gen/az/llama3.2-3b.sh | 7 ++++--- .../bash/experiment/rb_gen/az/test-azure.sh | 2 +- lib/style-transfer/configs/rb_gen/score/default.yaml | 8 ++++---- 5 files changed, 17 insertions(+), 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 index 4d64aca..feead6b 100644 --- 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 index 4c05844..a26b99b 100644 --- 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 index 78c2fab..0e67111 100644 --- 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 index b9592d7..ed569cb 100644 --- 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 index 6021d27..5806ae9 100644 --- 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