| pretraining_dataset: | |
| - path: bigcode/the-stack-v2 | |
| type: pretrain | |
| text_column: text | |
| split: train | |
| base_model: NousResearch/Llama-3.2-1B | |
| streaming_multipack_buffer_size: 10000 | |
| sample_packing: true | |
| sdp_attention: True | |
| adapter: lora | |
| adapter_r: 8 | |
| sequence_len: 512 | |
| train_batch_size: 1 | |
| warmup_ratio: 0.1 | |
| weight_decay: 0.0 | |
| learning_rate: 1e-4 | |
| max_steps: 2000 | |
| gradient_checkpointing: true |