Views
No views yet
Qwen/Qwen3.5-9B 训练的 LoRA adapter，用于判断两张图像中哪一张更接近任务完成状态。Qwen/Qwen3.5-9B1e-4checkpoint-2000adapter/img1 509 条，img2 491 条。| 模型或配置 | 最佳节点 | Accuracy | Balanced Accuracy | img1 Accuracy | img2 Accuracy | 无法解析 |
|---|---|---|---|---|---|---|
| Qwen3.5-9B 基模 | — | 53.90% | 54.64% | 13.36% | 95.93% | 1 |
LoRA，lr=1e-4 | 2000 | 79.00% | 79.01% | 78.59% | 79.43% | 0 |
LoRA，lr=5e-5 | 2500 | 77.60% | 77.57% | 79.37% | 75.76% | 0 |
img2 预测偏好，微调后两类准确率基本均衡。5e-5 实验作为学习率对照保留训练记录和评测汇总，但对应 adapter 未上传。| 参数 | 数值 |
|---|---|
| Base model | Qwen/Qwen3.5-9B |
| Tuner | LoRA |
| LoRA rank | 16 |
| LoRA alpha | 32 |
| LoRA dropout | 0.05 |
| Epochs | 2 |
| Train batch size | 4 |
| Eval batch size | 4 |
| Gradient accumulation | 2 |
| Effective batch size | 8 |
| Learning rate | 1e-4 / 5e-5 |
| LR scheduler | cosine |
| Warmup ratio | 0.05 |
| Precision | bfloat16 |
| Save/eval interval | 250 steps |
| PEFT version | 0.19.1 |
split_dataset_ratio=0，直接使用指定验证集1data/
2├── processed/
3│ ├── swift_train_10000.jsonl
4│ └── swift_val_1000.jsonl
5└── source_format/
6 ├── cot_trainset_10000.json
7 └── cot_trainset_val_1000.jsondata/processed/ 是训练脚本实际读取的 MS-SWIFT 格式；data/source_format/ 保存转换前的消息格式。/root/autodl-tmp/qwen35_spatiallogic/data/raw/SpatialLogic-Processed/...1.
2├── adapter/ # 最佳 LoRA adapter
3├── data/
4│ ├── processed/ # 实际训练输入
5│ └── source_format/ # 转换前数据
6├── results/
7│ ├── baseline/ # 基模评测结果
8│ ├── lr1e-4/ # 最佳实验及各节点汇总
9│ └── lr5e-5/ # 学习率对照汇总
10├── scripts/
11│ ├── data_preparation/ # 数据抽取与格式转换
12│ ├── train_lora_10000_2epochs_lr1e4.sh
13│ ├── train_lora_10000_2epochs_lr5e5.sh
14│ ├── infer_val.sh
15│ └── evaluate_img_pair.py
16└── training_logs/
17 ├── lr1e-4/
18 └── lr5e-5/1MODEL_PATH=/path/to/Qwen3.5-9B \
2bash scripts/infer_val.sh best_lora adapter1MODEL_PATH=/path/to/Qwen3.5-9B \
2bash scripts/infer_val.sh base_modelqwen35 的 Conda 环境。如果环境名称不同：1CONDA_ENV=your_env_name \
2MODEL_PATH=/path/to/Qwen3.5-9B \
3bash scripts/infer_val.sh best_lora adapter1python scripts/evaluate_img_pair.py \
2 --result-file outputs/infer_best_lora/val1000_bs32_TIMESTAMP.jsonl \
3 --val-file data/processed/swift_val_1000.jsonl \
4 --eval-dir outputs/eval_best_lora \
5 --run-name best_lora1e-4 配置：1MODEL_PATH=/path/to/Qwen3.5-9B \
2bash scripts/train_lora_10000_2epochs_lr1e4.sh5e-5 对照配置：1MODEL_PATH=/path/to/Qwen3.5-9B \
2bash scripts/train_lora_10000_2epochs_lr5e5.sh