warmup=0.1

configs/hotpot_qa.yaml

@@ -24,7 +24,7 @@ learning_rate: 0.00001
 # lr_scheduler_type: "constant_with_warmup"
 neftune_noise_alpha: 1
 weight_decay: 0.1
-warmup_ratio: 0.05
+warmup_ratio: 0.1
 
 # LoRA
 lora_r: 16

configs/pwc.yaml

@@ -24,7 +24,7 @@ learning_rate: 0.00001
 # lr_scheduler_type: "constant_with_warmup"
 neftune_noise_alpha: 1
 weight_decay: 0.1
-warmup_ratio: 0.05
+warmup_ratio: 0.1
 
 # LoRA
 lora_r: 16

configs/pwc_and_ctx_numbers_256.yaml

@@ -24,7 +24,7 @@ learning_rate: 0.00001
 # lr_scheduler_type: "constant_with_warmup"
 neftune_noise_alpha: 1
 weight_decay: 0.1
-warmup_ratio: 0.05
+warmup_ratio: 0.1
 
 # LoRA
 lora_r: 16

configs/pwc_hotpot_qa.yaml

@@ -0,0 +1,48 @@
+output_dir: "" # just a placeholder
+bf16: true
+model_name_or_path: meta-llama/Llama-3.2-1B-Instruct
+label_names: ["labels"]
+# eval_on_start: True
+# eval_strategy: "steps"
+# eval_steps: 500
+# save_strategy: "no"
+# # save_steps: 500
+# logging_strategy: "steps"
+# logging_steps: 100
+# use_liger_kernel: true
+# remove_unused_columns: false
+
+# needed to avoid OOM by compute the metrics batch by batch
+# w/o this the trainer stores logits of all sample in memory...
+# batch_eval_metrics: true
+
+per_device_train_batch_size: 32
+per_device_eval_batch_size: 32
+max_val_samples_per_ds: 1000
+# optim: schedule_free_adamw
+learning_rate: 0.00001
+# lr_scheduler_type: "constant_with_warmup"
+neftune_noise_alpha: 1
+weight_decay: 0.1
+warmup_ratio: 0.1
+
+# LoRA
+lora_r: 16
+lora_dropout: 0.05
+target_modules:
+  - down_proj
+  - up_proj
+  - gate_proj
+
+# data
+train_ds_names:
+- pwc
+- hotpot_qa
+
+val_ds_names:
+- pwc
+- hotpot_qa
+
+test_ds_names:
+- pwc
+- hotpot_qa