Create axolotl_config.yaml
Browse files- axolotl_config.yaml +45 -0
axolotl_config.yaml
ADDED
|
@@ -0,0 +1,45 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
base_model: Qwen/Qwen3-14B
|
| 2 |
+
|
| 3 |
+
plugins:
|
| 4 |
+
- axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
|
| 5 |
+
chat_template: qwen3
|
| 6 |
+
datasets:
|
| 7 |
+
- path: qingy2024/SynGen-Reasoning-V2
|
| 8 |
+
type: chat_template
|
| 9 |
+
split: train
|
| 10 |
+
field_messages: conversations
|
| 11 |
+
|
| 12 |
+
dataset_prepared_path: last_run_prepared
|
| 13 |
+
val_set_size: 0
|
| 14 |
+
output_dir: ./outputs/
|
| 15 |
+
|
| 16 |
+
sequence_len: 16384
|
| 17 |
+
sample_packing: true
|
| 18 |
+
|
| 19 |
+
wandb_project: huggingface
|
| 20 |
+
wandb_entity:
|
| 21 |
+
wandb_watch:
|
| 22 |
+
wandb_name: SynGen-14B
|
| 23 |
+
wandb_log_model:
|
| 24 |
+
|
| 25 |
+
gradient_accumulation_steps: 4
|
| 26 |
+
micro_batch_size: 8
|
| 27 |
+
num_epochs: 2
|
| 28 |
+
|
| 29 |
+
optimizer: adamw_torch_8bit
|
| 30 |
+
lr_scheduler: cosine
|
| 31 |
+
learning_rate: 2e-4
|
| 32 |
+
|
| 33 |
+
bf16: true
|
| 34 |
+
|
| 35 |
+
gradient_checkpointing: true
|
| 36 |
+
gradient_checkpointing_kwargs:
|
| 37 |
+
use_reentrant: false
|
| 38 |
+
flash_attention: true
|
| 39 |
+
|
| 40 |
+
logging_steps: 1
|
| 41 |
+
saves_per_epoch: 1
|
| 42 |
+
|
| 43 |
+
warmup_ratio: 0.1
|
| 44 |
+
weight_decay: 0.0
|
| 45 |
+
special_tokens:
|