Upload configs/sft_box_12g.yaml with huggingface_hub

configs/sft_box_12g.yaml

@@ -0,0 +1,50 @@
+# 专家 SFT (Box) - 适配 12G 显存
+# 从预训练 final checkpoint 继续，统一训练 grounding + counting + spatial
+# 使用统一 thinking 模板: "1. Analyzing → 2. Object grounding → 3. Conclusion"
+
+model_name_or_path: "outputs/pretrain/final"
+freeze_vision_tower: true
+freeze_llm: false
+use_lora: true
+lora_r: 64
+lora_alpha: 128
+lora_target_modules:
+  - "q_proj"
+  - "k_proj"
+  - "v_proj"
+  - "o_proj"
+  - "gate_proj"
+  - "up_proj"
+  - "down_proj"
+lora_dropout: 0.05
+torch_dtype: "bfloat16"
+
+load_in_4bit: false
+load_in_8bit: false
+
+datasets:
+  - path: "data/sft/grounding/sft_grounding.jsonl"
+    image_root: "data/coco/train/coco/train"
+    task_type: "grounding"
+    weight: 1.0
+  - path: "data/sft/counting/counting_data.jsonl"
+    image_root: "data/coco/train/coco/train"
+    task_type: "counting"
+    weight: 1.0
+  - path: "data/sft/spatial/spatial_data.jsonl"
+    image_root: "data/sft/spatial"
+    task_type: "spatial"
+    weight: 1.0
+
+total_samples: 40000
+batch_size: 1
+gradient_accumulation_steps: 16
+epochs: 5
+learning_rate: 5.0e-5
+weight_decay: 0.01
+warmup_ratio: 0.03
+max_length: 1024
+image_size: 336
+num_workers: 0
+save_every: 1
+device: "cuda"