Upload sft_llama_factory
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +1 -0
- sft_llama_factory/configs/deepseek_coder_67b_instruct_lora.yaml +43 -0
- sft_llama_factory/configs/deepseek_coder_67b_instruct_lora_r3lgw2.yaml +43 -0
- sft_llama_factory/configs/deepseek_coder_67b_instruct_lora_w0.yaml +43 -0
- sft_llama_factory/configs/deepseek_coder_67b_lora.yaml +43 -0
- sft_llama_factory/configs/deepseek_coder_67b_lora_r3lgw2.yaml +43 -0
- sft_llama_factory/configs/deepseek_coder_67b_lora_w0.yaml +43 -0
- sft_llama_factory/configs/qwen25_14b_instruct_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_14b_instruct_qlora_r3lgw2.yaml +47 -0
- sft_llama_factory/configs/qwen25_14b_instruct_qlora_w0.yaml +47 -0
- sft_llama_factory/configs/qwen25_14b_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_14b_qlora_r3lgw2.yaml +47 -0
- sft_llama_factory/configs/qwen25_3b_instruct_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_7b_instruct_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_coder_05b_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_coder_14b_instruct_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_coder_14b_instruct_qlora_r3lgw2.yaml +47 -0
- sft_llama_factory/configs/qwen25_coder_14b_instruct_qlora_w0.yaml +47 -0
- sft_llama_factory/configs/qwen25_coder_14b_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_coder_14b_qlora_r3lgw2.yaml +47 -0
- sft_llama_factory/configs/qwen25_coder_15b_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_coder_3b_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_coder_7b_instruct_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_coder_7b_lora.yaml +43 -0
- sft_llama_factory/configs/qwen25_math_7b_instruct_lora.yaml +43 -0
- sft_llama_factory/data/dataset_info.json +30 -0
- sft_llama_factory/data/minimal_edit_sft_mix_train.json +3 -0
- sft_llama_factory/data/minimal_edit_sft_nonood_test.json +0 -0
- sft_llama_factory/pids/deepseek_coder_67b_instruct_lora.pid +1 -0
- sft_llama_factory/pids/deepseek_coder_67b_instruct_lora_r3lgw2.pid +1 -0
- sft_llama_factory/pids/deepseek_coder_67b_instruct_lora_w0.pid +1 -0
- sft_llama_factory/pids/deepseek_coder_67b_lora.pid +1 -0
- sft_llama_factory/pids/deepseek_coder_67b_lora_r3lgw2.pid +1 -0
- sft_llama_factory/pids/deepseek_coder_67b_lora_w0.pid +1 -0
- sft_llama_factory/pids/qwen25_14b_instruct_lora.pid +1 -0
- sft_llama_factory/pids/qwen25_14b_instruct_qlora_r3lgw2.pid +1 -0
- sft_llama_factory/pids/qwen25_14b_instruct_qlora_w0.pid +1 -0
- sft_llama_factory/pids/qwen25_14b_lora.pid +1 -0
- sft_llama_factory/pids/qwen25_14b_qlora_r3lgw2.pid +1 -0
- sft_llama_factory/pids/qwen25_3b_instruct_lora.pid +1 -0
- sft_llama_factory/pids/qwen25_7b_instruct_lora.pid +1 -0
- sft_llama_factory/pids/qwen25_coder_05b_lora.pid +1 -0
- sft_llama_factory/pids/qwen25_coder_14b_instruct_lora.pid +1 -0
- sft_llama_factory/pids/qwen25_coder_14b_instruct_qlora_r3lgw2.pid +1 -0
- sft_llama_factory/pids/qwen25_coder_14b_instruct_qlora_w0.pid +1 -0
- sft_llama_factory/pids/qwen25_coder_14b_lora.pid +1 -0
- sft_llama_factory/pids/qwen25_coder_14b_qlora_r3lgw2.pid +1 -0
- sft_llama_factory/pids/qwen25_coder_15b_lora.pid +1 -0
- sft_llama_factory/pids/qwen25_coder_3b_lora.pid +1 -0
- sft_llama_factory/pids/qwen25_coder_7b_instruct_lora.pid +1 -0
.gitattributes
CHANGED
|
@@ -191,3 +191,4 @@ rl_runs/rl_main_ood_lev_8b_steps50_exact_rootpod3_clean_ulimit_20260522_092726/w
|
|
| 191 |
rl_runs/rl_main_ood_lev_8b_steps50_exact_rootpod3_overlap457_20260522_091312/run_default/wandb/offline-run-20260522_091322-s5c15ogp/run-s5c15ogp.wandb filter=lfs diff=lfs merge=lfs -text
|
| 192 |
rl_runs/rl_scaling_qwen3_8b_dense_ood_lev_step50_retry3_20260520_1233/run_default/wandb/offline-run-20260520_123314-rgc6xu2b/run-rgc6xu2b.wandb filter=lfs diff=lfs merge=lfs -text
|
| 193 |
rl_runs/rl_scaling_qwen3_8b_ood_lev_step50_20260520_093519/run_default/wandb/offline-run-20260520_093634-17k5n460/run-17k5n460.wandb filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 191 |
rl_runs/rl_main_ood_lev_8b_steps50_exact_rootpod3_overlap457_20260522_091312/run_default/wandb/offline-run-20260522_091322-s5c15ogp/run-s5c15ogp.wandb filter=lfs diff=lfs merge=lfs -text
|
| 192 |
rl_runs/rl_scaling_qwen3_8b_dense_ood_lev_step50_retry3_20260520_1233/run_default/wandb/offline-run-20260520_123314-rgc6xu2b/run-rgc6xu2b.wandb filter=lfs diff=lfs merge=lfs -text
|
| 193 |
rl_runs/rl_scaling_qwen3_8b_ood_lev_step50_20260520_093519/run_default/wandb/offline-run-20260520_093634-17k5n460/run-17k5n460.wandb filter=lfs diff=lfs merge=lfs -text
|
| 194 |
+
sft_llama_factory/data/minimal_edit_sft_mix_train.json filter=lfs diff=lfs merge=lfs -text
|
sft_llama_factory/configs/deepseek_coder_67b_instruct_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: deepseek-ai/deepseek-coder-6.7b-instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: deepseekcoder
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/deepseek_coder_67b_instruct_lora
|
sft_llama_factory/configs/deepseek_coder_67b_instruct_lora_r3lgw2.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: deepseek-ai/deepseek-coder-6.7b-instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: deepseekcoder
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/deepseek_coder_67b_instruct_lora_r3lgw2
|
sft_llama_factory/configs/deepseek_coder_67b_instruct_lora_w0.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: deepseek-ai/deepseek-coder-6.7b-instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: deepseekcoder
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/deepseek_coder_67b_instruct_lora_w0
|
sft_llama_factory/configs/deepseek_coder_67b_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: deepseek-ai/deepseek-coder-6.7b-base
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: deepseekcoder
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/deepseek_coder_67b_lora
|
sft_llama_factory/configs/deepseek_coder_67b_lora_r3lgw2.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: deepseek-ai/deepseek-coder-6.7b-base
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: deepseekcoder
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/deepseek_coder_67b_lora_r3lgw2
|
sft_llama_factory/configs/deepseek_coder_67b_lora_w0.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: deepseek-ai/deepseek-coder-6.7b-base
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: deepseekcoder
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/deepseek_coder_67b_lora_w0
|
sft_llama_factory/configs/qwen25_14b_instruct_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-14B-Instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_14b_instruct_lora
|
sft_llama_factory/configs/qwen25_14b_instruct_qlora_r3lgw2.yaml
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-14B-Instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
quantization_bit: 4
|
| 6 |
+
quantization_type: nf4
|
| 7 |
+
double_quantization: true
|
| 8 |
+
upcast_layernorm: true
|
| 9 |
+
stage: sft
|
| 10 |
+
do_train: true
|
| 11 |
+
finetuning_type: lora
|
| 12 |
+
lora_rank: 16
|
| 13 |
+
lora_alpha: 32
|
| 14 |
+
lora_dropout: 0.05
|
| 15 |
+
lora_target: all
|
| 16 |
+
|
| 17 |
+
### dataset
|
| 18 |
+
dataset: minimal_edit_sft_mix_train
|
| 19 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 20 |
+
template: qwen
|
| 21 |
+
cutoff_len: 4096
|
| 22 |
+
max_samples: 5000
|
| 23 |
+
overwrite_cache: true
|
| 24 |
+
preprocessing_num_workers: 8
|
| 25 |
+
val_size: 0.04
|
| 26 |
+
|
| 27 |
+
### output
|
| 28 |
+
logging_steps: 10
|
| 29 |
+
save_steps: 250
|
| 30 |
+
plot_loss: true
|
| 31 |
+
overwrite_output_dir: true
|
| 32 |
+
save_only_model: false
|
| 33 |
+
report_to: none
|
| 34 |
+
|
| 35 |
+
### train
|
| 36 |
+
per_device_train_batch_size: 1
|
| 37 |
+
gradient_accumulation_steps: 8
|
| 38 |
+
learning_rate: 2.0e-5
|
| 39 |
+
num_train_epochs: 1.0
|
| 40 |
+
lr_scheduler_type: cosine
|
| 41 |
+
warmup_ratio: 0.05
|
| 42 |
+
bf16: true
|
| 43 |
+
seed: 20260516
|
| 44 |
+
evaluation_strategy: steps
|
| 45 |
+
eval_steps: 125
|
| 46 |
+
per_device_eval_batch_size: 1
|
| 47 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_14b_instruct_qlora_r3lgw2
|
sft_llama_factory/configs/qwen25_14b_instruct_qlora_w0.yaml
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-14B-Instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
quantization_bit: 4
|
| 6 |
+
quantization_type: nf4
|
| 7 |
+
double_quantization: true
|
| 8 |
+
upcast_layernorm: true
|
| 9 |
+
stage: sft
|
| 10 |
+
do_train: true
|
| 11 |
+
finetuning_type: lora
|
| 12 |
+
lora_rank: 16
|
| 13 |
+
lora_alpha: 32
|
| 14 |
+
lora_dropout: 0.05
|
| 15 |
+
lora_target: all
|
| 16 |
+
|
| 17 |
+
### dataset
|
| 18 |
+
dataset: minimal_edit_sft_mix_train
|
| 19 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 20 |
+
template: qwen
|
| 21 |
+
cutoff_len: 4096
|
| 22 |
+
max_samples: 5000
|
| 23 |
+
overwrite_cache: true
|
| 24 |
+
preprocessing_num_workers: 8
|
| 25 |
+
val_size: 0.04
|
| 26 |
+
|
| 27 |
+
### output
|
| 28 |
+
logging_steps: 10
|
| 29 |
+
save_steps: 250
|
| 30 |
+
plot_loss: true
|
| 31 |
+
overwrite_output_dir: true
|
| 32 |
+
save_only_model: false
|
| 33 |
+
report_to: none
|
| 34 |
+
|
| 35 |
+
### train
|
| 36 |
+
per_device_train_batch_size: 1
|
| 37 |
+
gradient_accumulation_steps: 8
|
| 38 |
+
learning_rate: 2.0e-5
|
| 39 |
+
num_train_epochs: 1.0
|
| 40 |
+
lr_scheduler_type: cosine
|
| 41 |
+
warmup_ratio: 0.05
|
| 42 |
+
bf16: true
|
| 43 |
+
seed: 20260516
|
| 44 |
+
evaluation_strategy: steps
|
| 45 |
+
eval_steps: 125
|
| 46 |
+
per_device_eval_batch_size: 1
|
| 47 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_14b_instruct_qlora_w0
|
sft_llama_factory/configs/qwen25_14b_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-14B
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_14b_lora
|
sft_llama_factory/configs/qwen25_14b_qlora_r3lgw2.yaml
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-14B
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
quantization_bit: 4
|
| 6 |
+
quantization_type: nf4
|
| 7 |
+
double_quantization: true
|
| 8 |
+
upcast_layernorm: true
|
| 9 |
+
stage: sft
|
| 10 |
+
do_train: true
|
| 11 |
+
finetuning_type: lora
|
| 12 |
+
lora_rank: 16
|
| 13 |
+
lora_alpha: 32
|
| 14 |
+
lora_dropout: 0.05
|
| 15 |
+
lora_target: all
|
| 16 |
+
|
| 17 |
+
### dataset
|
| 18 |
+
dataset: minimal_edit_sft_mix_train
|
| 19 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 20 |
+
template: qwen
|
| 21 |
+
cutoff_len: 4096
|
| 22 |
+
max_samples: 5000
|
| 23 |
+
overwrite_cache: true
|
| 24 |
+
preprocessing_num_workers: 8
|
| 25 |
+
val_size: 0.04
|
| 26 |
+
|
| 27 |
+
### output
|
| 28 |
+
logging_steps: 10
|
| 29 |
+
save_steps: 250
|
| 30 |
+
plot_loss: true
|
| 31 |
+
overwrite_output_dir: true
|
| 32 |
+
save_only_model: false
|
| 33 |
+
report_to: none
|
| 34 |
+
|
| 35 |
+
### train
|
| 36 |
+
per_device_train_batch_size: 1
|
| 37 |
+
gradient_accumulation_steps: 8
|
| 38 |
+
learning_rate: 2.0e-5
|
| 39 |
+
num_train_epochs: 1.0
|
| 40 |
+
lr_scheduler_type: cosine
|
| 41 |
+
warmup_ratio: 0.05
|
| 42 |
+
bf16: true
|
| 43 |
+
seed: 20260516
|
| 44 |
+
evaluation_strategy: steps
|
| 45 |
+
eval_steps: 125
|
| 46 |
+
per_device_eval_batch_size: 1
|
| 47 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_14b_qlora_r3lgw2
|
sft_llama_factory/configs/qwen25_3b_instruct_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-3B-Instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_3b_instruct_lora
|
sft_llama_factory/configs/qwen25_7b_instruct_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-7B-Instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_7b_instruct_lora
|
sft_llama_factory/configs/qwen25_coder_05b_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Coder-0.5B
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_coder_05b_lora
|
sft_llama_factory/configs/qwen25_coder_14b_instruct_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Coder-14B-Instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_coder_14b_instruct_lora
|
sft_llama_factory/configs/qwen25_coder_14b_instruct_qlora_r3lgw2.yaml
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Coder-14B-Instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
quantization_bit: 4
|
| 6 |
+
quantization_type: nf4
|
| 7 |
+
double_quantization: true
|
| 8 |
+
upcast_layernorm: true
|
| 9 |
+
stage: sft
|
| 10 |
+
do_train: true
|
| 11 |
+
finetuning_type: lora
|
| 12 |
+
lora_rank: 16
|
| 13 |
+
lora_alpha: 32
|
| 14 |
+
lora_dropout: 0.05
|
| 15 |
+
lora_target: all
|
| 16 |
+
|
| 17 |
+
### dataset
|
| 18 |
+
dataset: minimal_edit_sft_mix_train
|
| 19 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 20 |
+
template: qwen
|
| 21 |
+
cutoff_len: 4096
|
| 22 |
+
max_samples: 5000
|
| 23 |
+
overwrite_cache: true
|
| 24 |
+
preprocessing_num_workers: 8
|
| 25 |
+
val_size: 0.04
|
| 26 |
+
|
| 27 |
+
### output
|
| 28 |
+
logging_steps: 10
|
| 29 |
+
save_steps: 250
|
| 30 |
+
plot_loss: true
|
| 31 |
+
overwrite_output_dir: true
|
| 32 |
+
save_only_model: false
|
| 33 |
+
report_to: none
|
| 34 |
+
|
| 35 |
+
### train
|
| 36 |
+
per_device_train_batch_size: 1
|
| 37 |
+
gradient_accumulation_steps: 8
|
| 38 |
+
learning_rate: 2.0e-5
|
| 39 |
+
num_train_epochs: 1.0
|
| 40 |
+
lr_scheduler_type: cosine
|
| 41 |
+
warmup_ratio: 0.05
|
| 42 |
+
bf16: true
|
| 43 |
+
seed: 20260516
|
| 44 |
+
evaluation_strategy: steps
|
| 45 |
+
eval_steps: 125
|
| 46 |
+
per_device_eval_batch_size: 1
|
| 47 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_coder_14b_instruct_qlora_r3lgw2
|
sft_llama_factory/configs/qwen25_coder_14b_instruct_qlora_w0.yaml
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Coder-14B-Instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
quantization_bit: 4
|
| 6 |
+
quantization_type: nf4
|
| 7 |
+
double_quantization: true
|
| 8 |
+
upcast_layernorm: true
|
| 9 |
+
stage: sft
|
| 10 |
+
do_train: true
|
| 11 |
+
finetuning_type: lora
|
| 12 |
+
lora_rank: 16
|
| 13 |
+
lora_alpha: 32
|
| 14 |
+
lora_dropout: 0.05
|
| 15 |
+
lora_target: all
|
| 16 |
+
|
| 17 |
+
### dataset
|
| 18 |
+
dataset: minimal_edit_sft_mix_train
|
| 19 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 20 |
+
template: qwen
|
| 21 |
+
cutoff_len: 4096
|
| 22 |
+
max_samples: 5000
|
| 23 |
+
overwrite_cache: true
|
| 24 |
+
preprocessing_num_workers: 8
|
| 25 |
+
val_size: 0.04
|
| 26 |
+
|
| 27 |
+
### output
|
| 28 |
+
logging_steps: 10
|
| 29 |
+
save_steps: 250
|
| 30 |
+
plot_loss: true
|
| 31 |
+
overwrite_output_dir: true
|
| 32 |
+
save_only_model: false
|
| 33 |
+
report_to: none
|
| 34 |
+
|
| 35 |
+
### train
|
| 36 |
+
per_device_train_batch_size: 1
|
| 37 |
+
gradient_accumulation_steps: 8
|
| 38 |
+
learning_rate: 2.0e-5
|
| 39 |
+
num_train_epochs: 1.0
|
| 40 |
+
lr_scheduler_type: cosine
|
| 41 |
+
warmup_ratio: 0.05
|
| 42 |
+
bf16: true
|
| 43 |
+
seed: 20260516
|
| 44 |
+
evaluation_strategy: steps
|
| 45 |
+
eval_steps: 125
|
| 46 |
+
per_device_eval_batch_size: 1
|
| 47 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_coder_14b_instruct_qlora_w0
|
sft_llama_factory/configs/qwen25_coder_14b_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Coder-14B
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_coder_14b_lora
|
sft_llama_factory/configs/qwen25_coder_14b_qlora_r3lgw2.yaml
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Coder-14B
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
quantization_bit: 4
|
| 6 |
+
quantization_type: nf4
|
| 7 |
+
double_quantization: true
|
| 8 |
+
upcast_layernorm: true
|
| 9 |
+
stage: sft
|
| 10 |
+
do_train: true
|
| 11 |
+
finetuning_type: lora
|
| 12 |
+
lora_rank: 16
|
| 13 |
+
lora_alpha: 32
|
| 14 |
+
lora_dropout: 0.05
|
| 15 |
+
lora_target: all
|
| 16 |
+
|
| 17 |
+
### dataset
|
| 18 |
+
dataset: minimal_edit_sft_mix_train
|
| 19 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 20 |
+
template: qwen
|
| 21 |
+
cutoff_len: 4096
|
| 22 |
+
max_samples: 5000
|
| 23 |
+
overwrite_cache: true
|
| 24 |
+
preprocessing_num_workers: 8
|
| 25 |
+
val_size: 0.04
|
| 26 |
+
|
| 27 |
+
### output
|
| 28 |
+
logging_steps: 10
|
| 29 |
+
save_steps: 250
|
| 30 |
+
plot_loss: true
|
| 31 |
+
overwrite_output_dir: true
|
| 32 |
+
save_only_model: false
|
| 33 |
+
report_to: none
|
| 34 |
+
|
| 35 |
+
### train
|
| 36 |
+
per_device_train_batch_size: 1
|
| 37 |
+
gradient_accumulation_steps: 8
|
| 38 |
+
learning_rate: 2.0e-5
|
| 39 |
+
num_train_epochs: 1.0
|
| 40 |
+
lr_scheduler_type: cosine
|
| 41 |
+
warmup_ratio: 0.05
|
| 42 |
+
bf16: true
|
| 43 |
+
seed: 20260516
|
| 44 |
+
evaluation_strategy: steps
|
| 45 |
+
eval_steps: 125
|
| 46 |
+
per_device_eval_batch_size: 1
|
| 47 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_coder_14b_qlora_r3lgw2
|
sft_llama_factory/configs/qwen25_coder_15b_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Coder-1.5B
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_coder_15b_lora
|
sft_llama_factory/configs/qwen25_coder_3b_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Coder-3B
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_coder_3b_lora
|
sft_llama_factory/configs/qwen25_coder_7b_instruct_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Coder-7B-Instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_coder_7b_instruct_lora
|
sft_llama_factory/configs/qwen25_coder_7b_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Coder-7B
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_coder_7b_lora
|
sft_llama_factory/configs/qwen25_math_7b_instruct_lora.yaml
ADDED
|
@@ -0,0 +1,43 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
### model
|
| 2 |
+
model_name_or_path: Qwen/Qwen2.5-Math-7B-Instruct
|
| 3 |
+
|
| 4 |
+
### method
|
| 5 |
+
stage: sft
|
| 6 |
+
do_train: true
|
| 7 |
+
finetuning_type: lora
|
| 8 |
+
lora_rank: 16
|
| 9 |
+
lora_alpha: 32
|
| 10 |
+
lora_dropout: 0.05
|
| 11 |
+
lora_target: all
|
| 12 |
+
|
| 13 |
+
### dataset
|
| 14 |
+
dataset: minimal_edit_sft_mix_train
|
| 15 |
+
dataset_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/data
|
| 16 |
+
template: qwen
|
| 17 |
+
cutoff_len: 4096
|
| 18 |
+
max_samples: 5000
|
| 19 |
+
overwrite_cache: true
|
| 20 |
+
preprocessing_num_workers: 8
|
| 21 |
+
val_size: 0.04
|
| 22 |
+
|
| 23 |
+
### output
|
| 24 |
+
logging_steps: 10
|
| 25 |
+
save_steps: 250
|
| 26 |
+
plot_loss: true
|
| 27 |
+
overwrite_output_dir: true
|
| 28 |
+
save_only_model: false
|
| 29 |
+
report_to: none
|
| 30 |
+
|
| 31 |
+
### train
|
| 32 |
+
per_device_train_batch_size: 1
|
| 33 |
+
gradient_accumulation_steps: 8
|
| 34 |
+
learning_rate: 2.0e-5
|
| 35 |
+
num_train_epochs: 1.0
|
| 36 |
+
lr_scheduler_type: cosine
|
| 37 |
+
warmup_ratio: 0.05
|
| 38 |
+
bf16: true
|
| 39 |
+
seed: 20260516
|
| 40 |
+
evaluation_strategy: steps
|
| 41 |
+
eval_steps: 125
|
| 42 |
+
per_device_eval_batch_size: 1
|
| 43 |
+
output_dir: /home/aiops/zhuty/fyp/artifacts/sft_llama_factory/saves/qwen25_math_7b_instruct_lora
|
sft_llama_factory/data/dataset_info.json
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"minimal_edit_sft_mix_train": {
|
| 3 |
+
"file_name": "minimal_edit_sft_mix_train.json",
|
| 4 |
+
"formatting": "sharegpt",
|
| 5 |
+
"columns": {
|
| 6 |
+
"messages": "conversations",
|
| 7 |
+
"system": "system"
|
| 8 |
+
},
|
| 9 |
+
"tags": {
|
| 10 |
+
"role_tag": "from",
|
| 11 |
+
"content_tag": "content",
|
| 12 |
+
"user_tag": "user",
|
| 13 |
+
"assistant_tag": "assistant"
|
| 14 |
+
}
|
| 15 |
+
},
|
| 16 |
+
"minimal_edit_sft_nonood_test": {
|
| 17 |
+
"file_name": "minimal_edit_sft_nonood_test.json",
|
| 18 |
+
"formatting": "sharegpt",
|
| 19 |
+
"columns": {
|
| 20 |
+
"messages": "conversations",
|
| 21 |
+
"system": "system"
|
| 22 |
+
},
|
| 23 |
+
"tags": {
|
| 24 |
+
"role_tag": "from",
|
| 25 |
+
"content_tag": "content",
|
| 26 |
+
"user_tag": "user",
|
| 27 |
+
"assistant_tag": "assistant"
|
| 28 |
+
}
|
| 29 |
+
}
|
| 30 |
+
}
|
sft_llama_factory/data/minimal_edit_sft_mix_train.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:918a4b0d37887d8072e7760dee14f58b9c8aef67fa20225f3a08f6e9174047e3
|
| 3 |
+
size 62127738
|
sft_llama_factory/data/minimal_edit_sft_nonood_test.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
sft_llama_factory/pids/deepseek_coder_67b_instruct_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
116785
|
sft_llama_factory/pids/deepseek_coder_67b_instruct_lora_r3lgw2.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
216740
|
sft_llama_factory/pids/deepseek_coder_67b_instruct_lora_w0.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
247382
|
sft_llama_factory/pids/deepseek_coder_67b_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
116774
|
sft_llama_factory/pids/deepseek_coder_67b_lora_r3lgw2.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
216730
|
sft_llama_factory/pids/deepseek_coder_67b_lora_w0.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
247372
|
sft_llama_factory/pids/qwen25_14b_instruct_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
116753
|
sft_llama_factory/pids/qwen25_14b_instruct_qlora_r3lgw2.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
216760
|
sft_llama_factory/pids/qwen25_14b_instruct_qlora_w0.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
247402
|
sft_llama_factory/pids/qwen25_14b_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
116763
|
sft_llama_factory/pids/qwen25_14b_qlora_r3lgw2.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
216780
|
sft_llama_factory/pids/qwen25_3b_instruct_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
380807
|
sft_llama_factory/pids/qwen25_7b_instruct_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
384607
|
sft_llama_factory/pids/qwen25_coder_05b_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
376034
|
sft_llama_factory/pids/qwen25_coder_14b_instruct_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
116743
|
sft_llama_factory/pids/qwen25_coder_14b_instruct_qlora_r3lgw2.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
216750
|
sft_llama_factory/pids/qwen25_coder_14b_instruct_qlora_w0.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
247392
|
sft_llama_factory/pids/qwen25_coder_14b_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
64733
|
sft_llama_factory/pids/qwen25_coder_14b_qlora_r3lgw2.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
216770
|
sft_llama_factory/pids/qwen25_coder_15b_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
376036
|
sft_llama_factory/pids/qwen25_coder_3b_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
380799
|
sft_llama_factory/pids/qwen25_coder_7b_instruct_lora.pid
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
383038
|