| base_model: mistralai/Mistral-7B-Instruct-v0.1 | |
| model_type: mistral | |
| tokenizer_type: mistral | |
| chat_template: alpaca | |
| datasets: | |
| - path: /mnt/data/warrungu_chat_dataset.json | |
| type: alpaca | |
| train_on_split: train | |
| output_dir: ./warrungu-test-output | |
| logging_dir: ./logs | |
| sequence_len: 512 | |
| sample_packing: false | |
| adapter: lora | |
| lora_r: 4 | |
| lora_alpha: 8 | |
| lora_dropout: 0.1 | |
| lora_target_linear: true | |
| gradient_accumulation_steps: 1 | |
| micro_batch_size: 1 | |
| batch_size: 1 # 👈 add this line to match batch sizing to data | |
| num_epochs: 1 | |
| learning_rate: 1e-4 | |
| lr_scheduler: cosine | |
| optimizer: adamw_bnb_8bit | |
| train_on_inputs: false | |
| group_by_length: false | |
| bf16: true | |
| tf32: true | |
| save_steps: 10 | |
| evals_per_epoch: 0 | |
| log_freq: 1 | |
| warmup_steps: 1 | |