gemma data exp

configs/pwc_hotpot_qa.yaml

@@ -0,0 +1,51 @@
+output_dir: "" # just a placeholder
+bf16: true
+model_name_or_path: google/gemma-2-2b-it
+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: 8
+per_device_eval_batch_size: 8
+max_val_samples_per_ds: 1000
+# optim: schedule_free_adamw
+
+learning_rate: 0.00004
+# lr_scheduler_type: "constant_with_warmup"
+neftune_noise_alpha: 1
+weight_decay: 0.01
+
+warmup_steps: 100
+
+dataloader_prefetch_factor: 8
+dataloader_num_workers: 8
+
+# LoRA
+lora_r: 8
+lora_dropout: 0.02
+target_modules:
+  - down_proj
+
+# data
+train_ds_names:
+  - pwc
+  - hotpot_qa
+
+val_ds_names:
+  - pwc
+  - hotpot_qa
+
+test_ds_names:
+  - pwc
+  - hotpot_qa

configs/qa.yaml

@@ -0,0 +1,65 @@
+output_dir: "" # just a placeholder
+bf16: true
+model_name_or_path: google/gemma-2-2b-it
+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: 8
+per_device_eval_batch_size: 8
+max_val_samples_per_ds: 1000
+# optim: schedule_free_adamw
+
+learning_rate: 0.00004
+# lr_scheduler_type: "constant_with_warmup"
+neftune_noise_alpha: 5
+weight_decay: 0.01
+#
+warmup_steps: 100
+
+dataloader_prefetch_factor: 8
+dataloader_num_workers: 8
+# LoRA
+lora_r: 8
+lora_dropout: 0.0
+target_modules:
+  - down_proj
+# data
+train_ds_names:
+  - fw_qa_3_small # ~ 20M
+  - ctx_qa # 300k
+  - pwc # 240k
+  - hotpot_qa # 90k
+  - squad # 90k
+  - drop # 77k
+  - narrativeqa # 40k
+  - quoref # 11k
+  - ropes # 11k
+  - synthetic_convqa # 40k
+
+val_ds_names:
+  - fw_qa_3_pretrain
+  - self_gen/google/gemma-2-2b-it/fw_qa_3
+  - self_gen/google/gemma-2-2b-it/ctx_qa
+  - self_gen/google/gemma-2-2b-it/pwc
+  - self_gen/google/gemma-2-2b-it/hotpot_qa
+  - self_gen/google/gemma-2-2b-it/squad
+  - fw_qa_3
+  - ctx_qa
+  - pwc
+  - hotpot_qa
+  - squad
+
+load_best_model_at_end: false
+metric_for_best_model: eval_pwc_loss

configs/self_gen_3_and_pretrain_small.yaml

@@ -0,0 +1,66 @@
+output_dir: "" # just a placeholder
+bf16: true
+model_name_or_path: google/gemma-2-2b-it
+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: 8
+per_device_eval_batch_size: 8
+max_val_samples_per_ds: 1000
+# optim: schedule_free_adamw
+
+learning_rate: 0.00004
+# lr_scheduler_type: "constant_with_warmup"
+neftune_noise_alpha: 5
+weight_decay: 0.01
+#
+warmup_steps: 100
+
+dataloader_prefetch_factor: 8
+dataloader_num_workers: 8
+# LoRA
+lora_r: 8
+lora_dropout: 0.0
+target_modules:
+  - down_proj
+# data
+train_ds_names:
+  - fw_qa_3_small_pretrain # ~20M
+  - self_gen/google/gemma-2-2b-it/fw_qa_3_small # ~20M
+  - self_gen/google/gemma-2-2b-it/ctx_qa # 300k
+  - self_gen/google/gemma-2-2b-it/pwc # 240k
+  - self_gen/google/gemma-2-2b-it/hotpot_qa # 90k
+  - self_gen/google/gemma-2-2b-it/squad # 90k
+  - self_gen/google/gemma-2-2b-it/drop # 77k
+  - self_gen/google/gemma-2-2b-it/narrativeqa # 40k
+  - self_gen/google/gemma-2-2b-it/quoref # 11k
+  - self_gen/google/gemma-2-2b-it/ropes # 11k
+  - self_gen/google/gemma-2-2b-it/synthetic_convqa # 40k
+
+val_ds_names:
+  - fw_qa_3_pretrain
+  - self_gen/google/gemma-2-2b-it/fw_qa_3
+  - self_gen/google/gemma-2-2b-it/ctx_qa
+  - self_gen/google/gemma-2-2b-it/pwc
+  - self_gen/google/gemma-2-2b-it/hotpot_qa
+  - self_gen/google/gemma-2-2b-it/squad
+  - fw_qa_3
+  - ctx_qa
+  - pwc
+  - hotpot_qa
+  - squad
+
+load_best_model_at_end: false
+metric_for_best_model: eval_pwc_loss

configs/self_gen_3_small.yaml

@@ -49,18 +49,17 @@ train_ds_names:
   - self_gen/google/gemma-2-2b-it/synthetic_convqa # 40k
 
 val_ds_names:
+  - fw_qa_3_pretrain
   - self_gen/google/gemma-2-2b-it/fw_qa_3
-  - self_gen/google/gemma-2-2b-it/fw_qa_xl
   - self_gen/google/gemma-2-2b-it/ctx_qa
   - self_gen/google/gemma-2-2b-it/pwc
   - self_gen/google/gemma-2-2b-it/hotpot_qa
   - self_gen/google/gemma-2-2b-it/squad
   - fw_qa_3
-  - fw_qa_xl
   - ctx_qa
   - pwc
   - hotpot_qa
   - squad
 
-load_best_model_at_end: true
+load_best_model_at_end: false
 metric_for_best_model: eval_pwc_loss

scripts/gemma_data_exp/gemma_qa.sh

@@ -0,0 +1,23 @@
+#!/bin/bash
+#SBATCH --job-name=ctxlora_medium
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29561 intx_sft.py configs/qa.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=4 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_self_attends_per_block=4 --num_latent_factor=1 \
+--lora_r=8 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True --per_rank_gen=True \
+--per_layer_processing=True \
+--gen_lora_l1_reg_coef=0.1 \
+

scripts/gemma_data_exp/gemma_self_gen_qa.sh

@@ -0,0 +1,23 @@
+#!/bin/bash
+#SBATCH --job-name=ctxlora_medium
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=8 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29562 intx_sft.py configs/self_gen_3_small.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=4 --per_device_train_batch_size=32 \
+--gradient_accumulation_steps=8 --per_device_eval_batch_size=32 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_self_attends_per_block=4 --num_latent_factor=1 \
+--lora_r=8 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True --per_rank_gen=True \
+--per_layer_processing=True \
+--gen_lora_l1_reg_coef=0.1 \
+

scripts/gemma_data_exp/gemma_self_gen_qa_and_pretrain.sh

@@ -0,0 +1,23 @@
+#!/bin/bash
+#SBATCH --job-name=ctxlora_medium
+#SBATCH --partition=a3
+#SBATCH --nodes=1
+#SBATCH --gpus=4
+#SBATCH --output=outputs/%x-%j.out
+#SBATCH --error=outputs/%x-%j.out
+
+uv run accelerate launch --num_processes=4 --gradient_accumulation_steps=32 --gradient_clipping=1.0 \
+--gpu_ids all --main_process_port 29563 intx_sft.py configs/self_gen_3_small_and_pretrain.yaml \
+--model_name_or_path=google/gemma-2-2b-it --num_train_epochs=2 --per_device_train_batch_size=8 \
+--gradient_accumulation_steps=32 --per_device_eval_batch_size=8 --exp_setup=hyper_lora --aggregator_type=perceiver \
+--target_modules=down_proj \
+--num_self_attends_per_block=4 --num_latent_factor=1 \
+--lora_r=8 \
+--eval_steps=5000 --save_steps=5000 --learning_rate=4e-5 --lora_dropout=0.0 \
+--neftune_noise_alpha=5 --use_light_weight_lora=False \
+--add_negative_prompt=False \
+--add_repeat_prompt=False \
+--use_sequence_packing=True --per_rank_gen=True \
+--per_layer_processing=True \
+--gen_lora_l1_reg_coef=0.1 \
+