| 1234567891011121314151617181920212223242526272829303132333435363738 |
- # SFT Training Config - v4 with /only: support
- # Usage: accelerate launch --config_file configs/accelerate_multi_gpu.yaml train.py sft --config configs/sft_v4.yaml
- model:
- base: "Qwen/Qwen3-1.7B"
- output: "qmd-sft-v4"
- dataset:
- name: "data/train_v4"
- text_field: "text"
- split: "train"
- eval_split: 0.1
- training:
- epochs: 3
- batch_size: 2
- gradient_accumulation_steps: 4
- learning_rate: 0.0002
- max_length: 512
- warmup_ratio: 0.03
- lr_scheduler: "cosine"
- lora:
- rank: 16
- alpha: 32
- dropout: 0.0
- target_modules:
- - "q_proj"
- - "k_proj"
- - "v_proj"
- - "o_proj"
- - "gate_proj"
- - "up_proj"
- - "down_proj"
- tracking:
- project: "qmd-query-expansion"
- run_name: "sft-1.7B-v4-only-modes"
|
