| # QLoRA for Qwen3-Coder-Next on Apple Silicon with 128 GB unified memory. | |
| model: mlx-community/Qwen3-Coder-Next-4bit | |
| train: true | |
| data: artifacts/bankai-v1/data | |
| fine_tune_type: lora | |
| optimizer: adamw | |
| mask_prompt: true | |
| num_layers: 16 | |
| batch_size: 1 | |
| grad_accumulation_steps: 2 | |
| iters: 612 | |
| val_batches: 8 | |
| learning_rate: 1.0e-5 | |
| steps_per_report: 5 | |
| steps_per_eval: 100 | |
| save_every: 100 | |
| adapter_path: artifacts/bankai-v1/adapter | |
| max_seq_length: 2048 | |
| grad_checkpoint: true | |
| clear_cache_threshold: 68719476736 | |
| seed: 42 | |
| lora_parameters: | |
| rank: 8 | |
| dropout: 0.0 | |
| scale: 16.0 | |
| keys: | |
| - linear_attn.in_proj_qkvz | |
| - linear_attn.in_proj_ba | |
| - linear_attn.out_proj | |
| - self_attn.q_proj | |
| - self_attn.k_proj | |
| - self_attn.v_proj | |
| - self_attn.o_proj | |