From e43b3f8aa36c198b95c1e46bea2eaf3893b13dc3 Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Wed, 17 Dec 2025 04:29:37 -0600 Subject: Initial commit (clean history) --- configs/qwen2.5_1.5b_full_sft.yaml | 33 +++++++++++++++++++++++++++++++++ 1 file changed, 33 insertions(+) create mode 100644 configs/qwen2.5_1.5b_full_sft.yaml (limited to 'configs/qwen2.5_1.5b_full_sft.yaml') diff --git a/configs/qwen2.5_1.5b_full_sft.yaml b/configs/qwen2.5_1.5b_full_sft.yaml new file mode 100644 index 0000000..e91176b --- /dev/null +++ b/configs/qwen2.5_1.5b_full_sft.yaml @@ -0,0 +1,33 @@ +### Qwen2.5-1.5B Full SFT Config +model_name_or_path: Qwen/Qwen2.5-1.5B-Instruct +stage: sft +do_train: true +finetuning_type: full + +dataset: preference_extractor_train +template: qwen +cutoff_len: 1024 +overwrite_cache: true +preprocessing_num_workers: 16 + +output_dir: saves/qwen2.5-1.5b-full-sft +logging_steps: 10 +save_strategy: steps +save_steps: 500 +plot_loss: true +overwrite_output_dir: true + +per_device_train_batch_size: 8 +gradient_accumulation_steps: 16 +learning_rate: 2.0e-5 +num_train_epochs: 1.0 +lr_scheduler_type: cosine +warmup_ratio: 0.05 +bf16: true +flash_attn: fa2 + +val_size: 0.01 +per_device_eval_batch_size: 8 +eval_strategy: steps +eval_steps: 500 + -- cgit v1.2.3