Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,9 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+checkpoints/rl-model/debug_log.txt filter=lfs diff=lfs merge=lfs -text
+checkpoints/rl-model/metrics.png filter=lfs diff=lfs merge=lfs -text
+checkpoints/rl-model/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+checkpoints/sft-model/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+dataset/sft/cot_filtered_sft.json filter=lfs diff=lfs merge=lfs -text
+eval/benchmarks/eval_ssv2_validation.json filter=lfs diff=lfs merge=lfs -text

checkpoints/rl-model/added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

checkpoints/rl-model/chat_template.json

@@ -0,0 +1,3 @@
+{
+  "chat_template": "{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n{% endif %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}<|im_end|>\n{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+}

checkpoints/rl-model/config.json

@@ -0,0 +1,50 @@
+{
+  "_name_or_path": "/root/zwj/Qwen2.5-VL-7B-sft-model",
+  "architectures": [
+    "Qwen2_5_VLForConditionalGeneration"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 3584,
+  "image_token_id": 151655,
+  "initializer_range": 0.02,
+  "intermediate_size": 18944,
+  "max_position_embeddings": 128000,
+  "max_window_layers": 28,
+  "model_type": "qwen2_5_vl",
+  "num_attention_heads": 28,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 4,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": {
+    "mrope_section": [
+      16,
+      24,
+      24
+    ],
+    "rope_type": "default",
+    "type": "default"
+  },
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.49.0.dev0",
+  "use_cache": false,
+  "use_sliding_window": false,
+  "video_token_id": 151656,
+  "vision_config": {
+    "hidden_size": 1280,
+    "in_chans": 3,
+    "model_type": "qwen2_5_vl",
+    "spatial_patch_size": 14,
+    "tokens_per_second": 2,
+    "torch_dtype": "float32"
+  },
+  "vision_end_token_id": 151653,
+  "vision_start_token_id": 151652,
+  "vision_token_id": 151654,
+  "vocab_size": 152064
+}

checkpoints/rl-model/debug_log.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a412f0f57e5b007d96a6a35607fdd6adb5c71b191251c8eff8c9cc91ad7ed1f
+size 142220165

checkpoints/rl-model/generation_config.json

@@ -0,0 +1,14 @@
+{
+  "attn_implementation": "flash_attention_2",
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.05,
+  "temperature": 1e-06,
+  "transformers_version": "4.49.0.dev0",
+  "use_cache": false
+}

checkpoints/rl-model/model-00001-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ba00fdb5573cbbbf0436e0b574c50686f9d96553ffff38bc11fdc21d5c5c82c
+size 4968243304

checkpoints/rl-model/model-00002-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f670ea93e36b0264abc861ed530ad96be420101e2866367c7cc69119dd852fe1
+size 4991495816

checkpoints/rl-model/model-00003-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a12eda7ecf268ef56796306df3c46e31fd837d297b84b3599ff807c6a036b48
+size 4932751040

checkpoints/rl-model/model-00004-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:389cf59b4f102a77772f1a24bd3ef4303e2c96a370c56498a3c9ff8253d21d0d
+size 1691924384

checkpoints/rl-model/model.safetensors.index.json

@@ -0,0 +1,736 @@
+{
+  "metadata": {
+    "total_size": 16584333312
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00004-of-00004.safetensors",
+    "model.embed_tokens.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.26.mlp.down_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.26.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.26.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.mlp.down_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.mlp.gate_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.mlp.up_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.k_proj.bias": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.k_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.o_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.q_proj.bias": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.q_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.v_proj.bias": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.v_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.5.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.5.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.norm.weight": "model-00004-of-00004.safetensors",
+    "visual.blocks.0.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.merger.ln_q.weight": "model-00001-of-00004.safetensors",
+    "visual.merger.mlp.0.bias": "model-00001-of-00004.safetensors",
+    "visual.merger.mlp.0.weight": "model-00001-of-00004.safetensors",
+    "visual.merger.mlp.2.bias": "model-00001-of-00004.safetensors",
+    "visual.merger.mlp.2.weight": "model-00001-of-00004.safetensors",
+    "visual.patch_embed.proj.weight": "model-00001-of-00004.safetensors"
+  }
+}

checkpoints/rl-model/preprocessor_config.json

@@ -0,0 +1,29 @@
+{
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_processor_type": "Qwen2_5_VLImageProcessor",
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "max_pixels": 501760,
+  "merge_size": 2,
+  "min_pixels": 3136,
+  "patch_size": 14,
+  "processor_class": "Qwen2_5_VLProcessor",
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "longest_edge": 12845056,
+    "shortest_edge": 3136
+  },
+  "temporal_patch_size": 2
+}

checkpoints/rl-model/special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoints/rl-model/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5eee858c5123a4279c3e1f7b81247343f356ac767940b2692a928ad929543214
+size 11422063

checkpoints/rl-model/tokenizer_config.json

@@ -0,0 +1,213 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n{% endif %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}<|im_end|>\n{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "max_length": null,
+  "model_max_length": 131072,
+  "pad_to_multiple_of": null,
+  "pad_token": "<|endoftext|>",
+  "pad_token_type_id": 0,
+  "padding_side": "right",
+  "processor_class": "Qwen2_5_VLProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoints/rl-model/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2ff2cb693b88eb056e5c876821692cb9f6ac31755dda99e9d4e956c657665d5
+size 8440

checkpoints/sft-model/added_tokens.json

@@ -0,0 +1,24 @@
+{
+  "</tool_call>": 151658,
+  "<tool_call>": 151657,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

checkpoints/sft-model/chat_template.json

@@ -0,0 +1,3 @@
+{
+  "chat_template": "{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n{% endif %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}<|im_end|>\n{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}"
+}

checkpoints/sft-model/config.json

@@ -0,0 +1,50 @@
+{
+  "_name_or_path": "/root/zwj/Model/Qwen2.5-VL-7B-Instruct",
+  "architectures": [
+    "Qwen2_5_VLForConditionalGeneration"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "eos_token_id": 151645,
+  "hidden_act": "silu",
+  "hidden_size": 3584,
+  "image_token_id": 151655,
+  "initializer_range": 0.02,
+  "intermediate_size": 18944,
+  "max_position_embeddings": 128000,
+  "max_window_layers": 28,
+  "model_type": "qwen2_5_vl",
+  "num_attention_heads": 28,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 4,
+  "rms_norm_eps": 1e-06,
+  "rope_scaling": {
+    "mrope_section": [
+      16,
+      24,
+      24
+    ],
+    "rope_type": "default",
+    "type": "default"
+  },
+  "rope_theta": 1000000.0,
+  "sliding_window": 32768,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.49.0.dev0",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "video_token_id": 151656,
+  "vision_config": {
+    "hidden_size": 1280,
+    "in_chans": 3,
+    "model_type": "qwen2_5_vl",
+    "spatial_patch_size": 14,
+    "tokens_per_second": 2,
+    "torch_dtype": "float32"
+  },
+  "vision_end_token_id": 151653,
+  "vision_start_token_id": 151652,
+  "vision_token_id": 151654,
+  "vocab_size": 152064
+}

checkpoints/sft-model/generation_config.json

@@ -0,0 +1,12 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": true,
+  "eos_token_id": [
+    151645,
+    151643
+  ],
+  "pad_token_id": 151643,
+  "repetition_penalty": 1.05,
+  "temperature": 1e-06,
+  "transformers_version": "4.49.0.dev0"
+}

checkpoints/sft-model/model-00001-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f46d9a59293a1f4d2ec11a0b73b6f0ba0dd7bc4f4a528fc3278a502339af9c2b
+size 4968243304

checkpoints/sft-model/model-00002-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a267bd8d35011f3fc273b2a7df667da0cdcb1d2c2eac6f3c4397b7a86284a630
+size 4991495816

checkpoints/sft-model/model-00003-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a63d0282c73caea663ffd78fa9a2f2233acc7deb6b5b4af5824de0c06e290c07
+size 4932751040

checkpoints/sft-model/model-00004-of-00004.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4b7d3370247e316d297ca31432922641905ef15e448ffa2de08fd78d972b29f
+size 1691924384

checkpoints/sft-model/model.safetensors.index.json

@@ -0,0 +1,736 @@
+{
+  "metadata": {
+    "total_size": 16584333312
+  },
+  "weight_map": {
+    "lm_head.weight": "model-00004-of-00004.safetensors",
+    "model.embed_tokens.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.0.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.1.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.10.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.10.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.11.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.12.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.13.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.14.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.15.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.16.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.16.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.17.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.17.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.18.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.19.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.2.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.2.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.20.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.20.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.21.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.22.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.23.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.24.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.input_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.mlp.down_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.post_attention_layernorm.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.25.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.input_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.26.mlp.down_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.26.mlp.gate_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.mlp.up_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.post_attention_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.26.self_attn.k_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.k_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.o_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.q_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.q_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.v_proj.bias": "model-00003-of-00004.safetensors",
+    "model.layers.26.self_attn.v_proj.weight": "model-00003-of-00004.safetensors",
+    "model.layers.27.input_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.mlp.down_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.mlp.gate_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.mlp.up_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.post_attention_layernorm.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.k_proj.bias": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.k_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.o_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.q_proj.bias": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.q_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.v_proj.bias": "model-00004-of-00004.safetensors",
+    "model.layers.27.self_attn.v_proj.weight": "model-00004-of-00004.safetensors",
+    "model.layers.3.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.3.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.input_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.post_attention_layernorm.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.4.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.5.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.5.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.5.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.5.self_attn.k_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.k_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.o_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.q_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.q_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.v_proj.bias": "model-00001-of-00004.safetensors",
+    "model.layers.5.self_attn.v_proj.weight": "model-00001-of-00004.safetensors",
+    "model.layers.6.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.6.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.7.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.8.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.input_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.down_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.gate_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.mlp.up_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.post_attention_layernorm.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.k_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.k_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.o_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.q_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.q_proj.weight": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.v_proj.bias": "model-00002-of-00004.safetensors",
+    "model.layers.9.self_attn.v_proj.weight": "model-00002-of-00004.safetensors",
+    "model.norm.weight": "model-00004-of-00004.safetensors",
+    "visual.blocks.0.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.0.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.1.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.10.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.11.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.12.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.13.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.14.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.15.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.16.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.17.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.18.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.19.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.2.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.20.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.21.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.22.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.23.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.24.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.25.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.26.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.27.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.28.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.29.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.3.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.30.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.31.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.4.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.5.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.6.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.7.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.8.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.attn.proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.attn.proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.attn.qkv.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.attn.qkv.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.down_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.down_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.gate_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.gate_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.up_proj.bias": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.mlp.up_proj.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.norm1.weight": "model-00001-of-00004.safetensors",
+    "visual.blocks.9.norm2.weight": "model-00001-of-00004.safetensors",
+    "visual.merger.ln_q.weight": "model-00001-of-00004.safetensors",
+    "visual.merger.mlp.0.bias": "model-00001-of-00004.safetensors",
+    "visual.merger.mlp.0.weight": "model-00001-of-00004.safetensors",
+    "visual.merger.mlp.2.bias": "model-00001-of-00004.safetensors",
+    "visual.merger.mlp.2.weight": "model-00001-of-00004.safetensors",
+    "visual.patch_embed.proj.weight": "model-00001-of-00004.safetensors"
+  }
+}

checkpoints/sft-model/preprocessor_config.json

@@ -0,0 +1,29 @@
+{
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_processor_type": "Qwen2_5_VLImageProcessor",
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "max_pixels": 12845056,
+  "merge_size": 2,
+  "min_pixels": 3136,
+  "patch_size": 14,
+  "processor_class": "Qwen2_5_VLProcessor",
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "longest_edge": 12845056,
+    "shortest_edge": 3136
+  },
+  "temporal_patch_size": 2
+}

checkpoints/sft-model/special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

checkpoints/sft-model/tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba0c439f7be467bf47d12a7e6f9adc6116201056fc60c67f431c679b7c16afc8
+size 11422064

checkpoints/sft-model/tokenizer_config.json

@@ -0,0 +1,209 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% for message in messages %}{% if loop.first and message['role'] != 'system' %}<|im_start|>system\nYou are a helpful assistant.<|im_end|>\n{% endif %}<|im_start|>{{ message['role'] }}\n{% if message['content'] is string %}{{ message['content'] }}<|im_end|>\n{% else %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% if add_vision_id %}Picture {{ image_count.value }}: {% endif %}<|vision_start|><|image_pad|><|vision_end|>{% elif content['type'] == 'video' or 'video' in content %}{% set video_count.value = video_count.value + 1 %}{% if add_vision_id %}Video {{ video_count.value }}: {% endif %}<|vision_start|><|video_pad|><|vision_end|>{% elif 'text' in content %}{{ content['text'] }}{% endif %}{% endfor %}<|im_end|>\n{% endif %}{% endfor %}{% if add_generation_prompt %}<|im_start|>assistant\n{% endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "processor_class": "Qwen2_5_VLProcessor",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

checkpoints/sft-model/training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1028ce87a560fb384c42d9a60b025a8fefe81db78fb79f38c90fbe7dc204b838
+size 7864

cot_generate/add_timestamp_to_video.py

@@ -0,0 +1,157 @@
+"""
+给视频的每一帧左上角打上该帧对应的时间戳。
+时间戳格式: HH:MM:SS
+
+用法:
+    python add_timestamp_to_video.py --input video.mp4 --output video_with_timestamp.mp4
+    python add_timestamp_to_video.py --input video.mp4  # 默认输出到同目录下 xxx_timestamped.mp4
+    python add_timestamp_to_video.py --input_dir /path/to/videos --output_dir /path/to/output  # 批量处理
+"""
+
+import argparse
+import os
+import subprocess
+import tempfile
+import cv2
+import glob
+from pathlib import Path
+
+
+def format_timestamp(seconds: float) -> str:
+    """将秒数格式化为 HH:MM:SS"""
+    h = int(seconds // 3600)
+    m = int((seconds % 3600) // 60)
+    s = int(seconds % 60)
+    return f"{h:02d}:{m:02d}:{s:02d}"
+
+
+def add_timestamp_to_video(input_path: str, output_path: str, font_scale: float = 1.0, thickness: int = 2):
+    """
+    给视频每一帧的左上角添加时间戳。
+
+    Args:
+        input_path: 输入视频路径
+        output_path: 输出视频路径
+        font_scale: 字体大小缩放系数
+        thickness: 字体粗细
+    """
+    cap = cv2.VideoCapture(input_path)
+    if not cap.isOpened():
+        print(f"[ERROR] 无法打开视频: {input_path}")
+        return False
+
+    fps = cap.get(cv2.CAP_PROP_FPS)
+    width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+
+    # 根据视频分辨率自适应字体大小
+    auto_font_scale = font_scale * min(width, height) / 480.0
+    auto_thickness = max(1, int(thickness * min(width, height) / 480.0))
+
+    os.makedirs(os.path.dirname(output_path) if os.path.dirname(output_path) else ".", exist_ok=True)
+
+    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+    writer = cv2.VideoWriter(output_path, fourcc, fps, (width, height))
+
+    frame_idx = 0
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break
+
+        current_time = frame_idx / fps
+        timestamp_text = format_timestamp(current_time)
+
+        # 计算文字尺寸
+        font = cv2.FONT_HERSHEY_SIMPLEX
+        (text_w, text_h), baseline = cv2.getTextSize(timestamp_text, font, auto_font_scale, auto_thickness)
+
+        # 左上角位置，留一些边距
+        margin = int(10 * min(width, height) / 480.0)
+        x, y = margin, margin + text_h
+
+        # 画半透明黑色背景框，提高可读性
+        pad = int(6 * min(width, height) / 480.0)
+        overlay = frame.copy()
+        cv2.rectangle(overlay,
+                      (x - pad, y - text_h - pad),
+                      (x + text_w + pad, y + baseline + pad),
+                      (0, 0, 0), -1)
+        cv2.addWeighted(overlay, 0.6, frame, 0.4, 0, frame)
+
+        # 画白色时间戳文字
+        cv2.putText(frame, timestamp_text, (x, y),
+                    font, auto_font_scale, (255, 255, 255), auto_thickness, cv2.LINE_AA)
+
+        writer.write(frame)
+        frame_idx += 1
+
+        if frame_idx % 100 == 0:
+            print(f"  处理进度: {frame_idx}/{total_frames} ({frame_idx * 100 // total_frames}%)")
+
+    cap.release()
+    writer.release()
+
+    # 用 ffmpeg 重新编码为 H.264，大幅减小文件体积
+    tmp_path = output_path + ".tmp.mp4"
+    os.rename(output_path, tmp_path)
+    try:
+        cmd = [
+            "ffmpeg", "-y", "-i", tmp_path,
+            "-c:v", "libx264", "-preset", "medium", "-crf", "23",
+            "-pix_fmt", "yuv420p", "-an",
+            output_path
+        ]
+        subprocess.run(cmd, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL, check=True)
+        os.remove(tmp_path)
+    except (subprocess.CalledProcessError, FileNotFoundError) as e:
+        print(f"[WARN] ffmpeg 压缩失败({e})，使用未压缩版本")
+        if os.path.exists(tmp_path):
+            os.rename(tmp_path, output_path)
+
+    print(f"[DONE] {input_path} -> {output_path}  (共 {frame_idx} 帧)")
+    return True
+
+
+def main():
+    parser = argparse.ArgumentParser(description="给视频左上角打上帧时间戳")
+    parser.add_argument("--input", type=str, default=None, help="输入视频路径")
+    parser.add_argument("--output", type=str, default=None, help="输出视频路径 (默认在原文件名后加 _timestamped)")
+    parser.add_argument("--input_dir", type=str, default=None, help="批量处理: 输入视频文件夹")
+    parser.add_argument("--output_dir", type=str, default=None, help="批量处理: 输出视频文件夹")
+    parser.add_argument("--font_scale", type=float, default=1.0, help="字体大小缩放 (默认 1.0)")
+    parser.add_argument("--thickness", type=int, default=2, help="字体粗细 (默认 2)")
+    parser.add_argument("--ext", type=str, default="mp4", help="批量模式下匹配的视频扩展名 (默认 mp4)")
+    args = parser.parse_args()
+
+    if args.input:
+        # 单个视频处理
+        if args.output is None:
+            stem = Path(args.input).stem
+            suffix = Path(args.input).suffix
+            args.output = str(Path(args.input).parent / f"{stem}_timestamped{suffix}")
+        add_timestamp_to_video(args.input, args.output, args.font_scale, args.thickness)
+
+    elif args.input_dir:
+        # 批量处理
+        if args.output_dir is None:
+            args.output_dir = args.input_dir + "_timestamped"
+        os.makedirs(args.output_dir, exist_ok=True)
+
+        video_files = sorted(glob.glob(os.path.join(args.input_dir, f"*.{args.ext}")))
+        if not video_files:
+            print(f"[WARN] 在 {args.input_dir} 中未找到 *.{args.ext} 文件")
+            return
+
+        print(f"共找到 {len(video_files)} 个视频文件")
+        for i, vpath in enumerate(video_files, 1):
+            out_path = os.path.join(args.output_dir, os.path.basename(vpath))
+            print(f"\n[{i}/{len(video_files)}] 处理: {vpath}")
+            add_timestamp_to_video(vpath, out_path, args.font_scale, args.thickness)
+    else:
+        parser.print_help()
+
+
+if __name__ == "__main__":
+    main()

cot_generate/batch_add_timestamp.py

@@ -0,0 +1,118 @@
+"""
+批量给视频加时间戳水印 — 从 Video-R1 JSON 提取唯一视频路径,
+保持原始目录结构输出到新目录, 支持多进程并行和断点续跑。
+
+用法:
+    python batch_add_timestamp.py \
+        --input_json /path/to/Video-R1-260k.json \
+        --video_root /path/to/video_root \
+        --output_root /path/to/output_root \
+        --workers 8
+"""
+
+import argparse
+import json
+import os
+from pathlib import Path
+from multiprocessing import Pool
+from functools import partial
+
+from add_timestamp_to_video import add_timestamp_to_video
+
+
+def extract_unique_videos(json_path: str) -> list:
+    """从 Video-R1 JSON 中提取所有唯一的视频相对路径"""
+    with open(json_path, encoding="utf-8") as f:
+        data = json.load(f)
+    videos = set()
+    for item in data:
+        rel = item.get("path", "").lstrip("./")
+        if rel:
+            videos.add(rel)
+    return sorted(videos)
+
+
+def process_one(rel_path: str, video_root: str, output_root: str) -> dict:
+    """处理单个视频: 加时间戳水印"""
+    src = os.path.join(video_root, rel_path)
+    dst = os.path.join(output_root, rel_path)
+
+    if not os.path.exists(src):
+        return {"path": rel_path, "status": "missing", "msg": f"Source not found: {src}"}
+
+    if os.path.exists(dst):
+        return {"path": rel_path, "status": "skipped", "msg": "Already exists"}
+
+    try:
+        ok = add_timestamp_to_video(src, dst)
+        if ok:
+            return {"path": rel_path, "status": "done"}
+        else:
+            return {"path": rel_path, "status": "error", "msg": "add_timestamp returned False"}
+    except Exception as e:
+        return {"path": rel_path, "status": "error", "msg": str(e)}
+
+
+def main():
+    parser = argparse.ArgumentParser(description="批量加时间戳水印 (保持目录结构)")
+    parser.add_argument("--input_json", required=True,
+                        help="Video-R1 格式的 JSON 文件路径")
+    parser.add_argument("--video_root", required=True,
+                        help="原始视频的根目录")
+    parser.add_argument("--output_root", required=True,
+                        help="输出水印视频的根目录 (保持原始子目录结构)")
+    parser.add_argument("--workers", type=int, default=4,
+                        help="并行处理进程数 (默认 4)")
+    args = parser.parse_args()
+
+    videos = extract_unique_videos(args.input_json)
+    print(f"[INFO] Found {len(videos)} unique videos in {args.input_json}")
+
+    # 统计已完成的
+    already_done = sum(1 for v in videos if os.path.exists(os.path.join(args.output_root, v)))
+    if already_done > 0:
+        print(f"[INFO] {already_done} already timestamped, {len(videos) - already_done} remaining")
+
+    os.makedirs(args.output_root, exist_ok=True)
+
+    worker_fn = partial(process_one,
+                        video_root=args.video_root,
+                        output_root=args.output_root)
+
+    done = 0
+    errors = 0
+    skipped = 0
+
+    if args.workers <= 1:
+        for i, v in enumerate(videos, 1):
+            result = worker_fn(v)
+            if result["status"] == "done":
+                done += 1
+            elif result["status"] == "skipped":
+                skipped += 1
+            else:
+                errors += 1
+                print(f"  [ERROR] {result['path']}: {result.get('msg', '')}")
+            if i % 50 == 0:
+                print(f"  Progress: {i}/{len(videos)} (done={done}, skipped={skipped}, errors={errors})")
+    else:
+        with Pool(processes=args.workers) as pool:
+            for i, result in enumerate(pool.imap_unordered(worker_fn, videos), 1):
+                if result["status"] == "done":
+                    done += 1
+                elif result["status"] == "skipped":
+                    skipped += 1
+                else:
+                    errors += 1
+                    print(f"  [ERROR] {result['path']}: {result.get('msg', '')}")
+                if i % 50 == 0:
+                    print(f"  Progress: {i}/{len(videos)} (done={done}, skipped={skipped}, errors={errors})")
+
+    total = done + skipped
+    print(f"\n[DONE] Processed: {done} new, {skipped} skipped, {errors} errors")
+    print(f"[DONE] Total timestamped: {total}/{len(videos)}")
+    print(f"[DONE] Output: {args.output_root}")
+
+
+if __name__ == "__main__":
+    main()

cot_generate/clean_and_convert.py

@@ -0,0 +1,181 @@
+"""
+清洗 CoT 数据 + 转换为 Video-R1 SFT 格式
+
+输入: video_qa_final.jsonl (原始格式: video, question, answer, problem_id, response)
+输出: video_qa_sft.json (Video-R1 格式: problem_id, problem, data_type, problem_type,
+      options, process, solution, path, data_source)
+"""
+
+import re
+import json
+import argparse
+from pathlib import Path
+from clean_cot_data import clean_response, validate_item
+
+
+# 根据视频路径推断 data_source
+DATA_SOURCE_MAP = {
+    "activitynet": "ActivityNet",
+    "Charades": "Charades",
+    "youcook2": "YouCook2",
+    "clevrer": "CLEVRER",
+    "CLEVRER": "CLEVRER",
+    "chiral_actions": "Chiral_Actions",
+    "speed_actions": "Speed_Actions",
+    "youtube_video_2024": "LLaVA-Video-178K/youtube_video_2024",
+    "NextQA": "NeXT-QA",
+    "NExTVideo": "NeXT-QA",
+    "STAR": "STAR",
+    "PerceptionTest": "PerceptionTest",
+    "ego4d": "Ego4D",
+}
+
+
+def infer_data_source(video_path: str) -> str:
+    """根据视频路径推断 data_source"""
+    for key, source in DATA_SOURCE_MAP.items():
+        if f"/{key}/" in video_path or video_path.endswith(f"/{key}"):
+            return source
+    # 数字目录 → NeXT-QA
+    parts = video_path.split("/")
+    for i, p in enumerate(parts):
+        if p == "NExTVideo" or p == "NextQA":
+            return "NeXT-QA"
+    # fallback: 用路径中的关键目录
+    if "LLaVA-Video-178K" in video_path:
+        return "LLaVA-Video-178K"
+    if "NeXT-QA" in video_path:
+        return "NeXT-QA"
+    return ""
+
+
+def convert_path(video_path: str) -> str:
+    """将绝对路径转为相对路径 (相对于 dataset_First_Version)"""
+    marker = "/dataset_First_Version/"
+    if marker in video_path:
+        return "./" + video_path.split(marker, 1)[1]
+    return video_path
+
+
+def detect_problem_type(question: str, answer: str) -> tuple[str, list]:
+    """检测题目类型和选项"""
+    # 检查 answer 是否是选择题格式 (A/B/C/D 或 A. xxx)
+    if re.match(r'^[A-D]\.?\s', answer) or re.match(r'^[A-D]$', answer.strip()):
+        return "multiple choice", []
+    return "free-form", []
+
+
+def main():
+    parser = argparse.ArgumentParser(description="清洗并转换为 Video-R1 SFT 格式")
+    parser.add_argument("--input", required=True, help="输入 JSONL 文件")
+    parser.add_argument("--output", required=True, help="输出 JSON 文件")
+    parser.add_argument("--filter-wrong", action="store_true",
+                        help="过滤答案与GT不一致的条目")
+    args = parser.parse_args()
+
+    # 读取
+    items = []
+    with open(args.input, encoding="utf-8") as f:
+        for line in f:
+            if line.strip():
+                items.append(json.loads(line))
+    print(f"读取 {len(items)} 条数据")
+
+    # 清洗 + 转换
+    results = []
+    total_fixes = 0
+    total_warnings = 0
+    n_filtered = 0
+
+    for i, item in enumerate(items):
+        pid = item.get("problem_id", str(i))
+        video = item.get("video", "")
+        question = item.get("question", "")
+        answer = item.get("answer", "")
+        response = item.get("response", "")
+
+        if not response:
+            total_warnings += 1
+            print(f"  [{pid}] ⚠ response为空，跳过")
+            continue
+
+        # 1. 清洗 response
+        cleaned_resp, fixes = clean_response(response, answer)
+        if fixes:
+            total_fixes += len(fixes)
+
+        # 2. 验证
+        item["response"] = cleaned_resp
+        warnings = validate_item(item)
+        if warnings:
+            total_warnings += len(warnings)
+
+        # 3. 过滤答案错误
+        if args.filter_wrong:
+            ans_m = re.search(r"<answer>(.*?)</answer>", cleaned_resp, re.DOTALL)
+            model_ans = ans_m.group(1).strip() if ans_m else ""
+            if answer and model_ans != answer:
+                n_filtered += 1
+                continue
+
+        # 4. 拆分 process 和 solution
+        # process = <think>...</think> 部分
+        # solution = <answer>...</answer> 部分
+        think_match = re.search(r"(<think>.*?</think>)", cleaned_resp, re.DOTALL)
+        process = think_match.group(1) if think_match else cleaned_resp
+
+        answer_match = re.search(r"(<answer>.*?</answer>)", cleaned_resp, re.DOTALL)
+        solution = answer_match.group(0) if answer_match else f"<answer>{answer}</answer>"
+
+        # 如果 think 之后、answer 之前还有内容（如 Reference frame），也放入 process
+        # 完整的 process = think 外壳 + 所有分析内容
+        # 从 cleaned_resp 中去掉 answer 和 key_frame，剩下的都是 process
+        process_text = cleaned_resp
+        process_text = re.sub(r"<answer>.*?</answer>", "", process_text, flags=re.DOTALL)
+        process_text = re.sub(r"<key_frame>.*?</key_frame>", "", process_text, flags=re.DOTALL)
+        process_text = process_text.strip()
+        if process_text:
+            process = process_text
+
+        # 5. 推断类型
+        problem_type, options = detect_problem_type(question, answer)
+
+        # 6. 构建 Video-R1 格式
+        result = {
+            "problem_id": pid,
+            "problem": question,
+            "data_type": "video",
+            "problem_type": problem_type,
+            "options": options,
+            "process": process,
+            "solution": solution,
+            "path": convert_path(video),
+            "data_source": infer_data_source(video),
+        }
+        results.append(result)
+
+    # 写出 JSON 数组
+    with open(args.output, "w", encoding="utf-8") as f:
+        json.dump(results, f, ensure_ascii=False, indent=2)
+
+    # 报告
+    print(f"\n{'='*50}")
+    print(f"清洗并转换完成:")
+    print(f"  输入条数:   {len(items)}")
+    print(f"  输出条数:   {len(results)}")
+    print(f"  修复数:     {total_fixes}")
+    print(f"  警告数:     {total_warnings}")
+    if args.filter_wrong:
+        print(f"  过滤错误:   {n_filtered}")
+    print(f"  输出文件:   {args.output}")
+
+    # data_source 分布
+    from collections import Counter
+    src_dist = Counter(r["data_source"] for r in results)
+    print(f"\ndata_source 分布:")
+    for k, v in src_dist.most_common():
+        print(f"  {k}: {v}")
+
+
+if __name__ == "__main__":
+    main()

cot_generate/clean_cot_data.py

@@ -0,0 +1,296 @@
+"""
+CoT 数据清洗脚本
+
+功能:
+1. 自动识别 think 模式新数据 / 旧数据两种格式
+   - 新数据 (think 模式): <think>[草稿]</think> Ref-Delta <answer>
+     → 丢弃草稿，将 Ref-Delta 包入 <think>
+   - 旧数据: <think>[Ref-Delta]</think> <answer>  (保持原逻辑)
+2. 处理重复 <think></think> 标签
+3. 验证必要字段完整性
+4. 输出清洗报告
+
+用法:
+    python clean_cot_data.py input.jsonl -o output.jsonl
+    python clean_cot_data.py input.jsonl                    # 覆盖原文件
+"""
+
+import re
+import sys
+import json
+import argparse
+from pathlib import Path
+
+
+def clean_response(response: str, gt_answer: str = "") -> tuple[str, list[str]]:
+    """清洗单条 response，返回 (清洗后文本, 修复记录列表)"""
+    fixes = []
+
+    # ---- 1. 保护 <answer> 和 <key_frame> 块 ----
+    answer_match = re.search(r"<answer>(.*?)</answer>", response, re.DOTALL)
+    answer_block = answer_match.group(0).strip() if answer_match else ""
+
+    kf_match = re.search(r"<key_frame>(.*?)</key_frame>", response, re.DOTALL)
+    kf_block = kf_match.group(0).strip() if kf_match else ""
+
+    # ---- 2. 识别格式：新(think模式) vs 旧 ----
+    # 新格式特征：Ref-Delta 在 </think> 之后（草稿在 <think> 内部）
+    # 旧格式特征：Ref-Delta 在 <think> 内部
+    think_match = re.search(r"<think>(.*?)</think>(.*)", response, re.DOTALL)
+    is_new_format = False
+    if think_match:
+        after_think = think_match.group(2)
+        if "Reference frame at" in after_think or "<time_range>" in after_think:
+            is_new_format = True
+
+    if is_new_format:
+        # 新格式：丢弃 <think> 内的草稿，取 </think> 之后的 Ref-Delta 作为内容
+        after_think = think_match.group(2)
+        content = re.sub(r"<answer>.*?</answer>", "", after_think, flags=re.DOTALL)
+        content = re.sub(r"<key_frame>.*?</key_frame>", "", content, flags=re.DOTALL)
+        content = content.strip()
+        fixes.append("think模式新格式: 去除内部草稿，保留Ref-Delta")
+        n_think_open = n_think_close = 1
+    else:
+        # 旧格式：去掉所有 <think></think> 标签，内容原样保留
+        body = response
+        body = re.sub(r"<answer>.*?</answer>", "", body, flags=re.DOTALL)
+        body = re.sub(r"<key_frame>.*?</key_frame>", "", body, flags=re.DOTALL)
+        n_think_open = len(re.findall(r"<think>", body))
+        n_think_close = len(re.findall(r"</think>", body))
+        content = re.sub(r"</?think>", "", body).strip()
+
+    # ---- 3. 清理 think 内容 ----
+    # 3a. 把 bullet list 转为散文 (去掉行首 - • * 符号)
+    if re.search(r"^\s*[-•*]\s", content, re.MULTILINE):
+        lines = content.split("\n")
+        cleaned_lines = []
+        for line in lines:
+            stripped = re.sub(r"^\s*[-•*]\s+", "", line)
+            if stripped != line:
+                fixes.append(f"去除bullet: {line.strip()[:60]}")
+            cleaned_lines.append(stripped)
+        content = "\n".join(cleaned_lines)
+
+    # 3b. 把 "At <time>..." 句式统一为 Reference frame 句式
+    def _fix_at_time(m):
+        full = m.group(0)
+        t = m.group(1)
+        rest = m.group(2).strip().rstrip(".")
+        fixes.append(f"At <time> → Reference frame: {full[:60]}")
+        return f"Reference frame at <time>{t}</time>: {rest}."
+    content = re.sub(
+        r"(?<!Reference frame )At <time>(.*?)</time>[,:]\s*(.*?)(?=\.\s|\n\n|$)",
+        _fix_at_time, content, flags=re.DOTALL,
+    )
+
+    # ---- 4. 记录修复 ----
+    if not is_new_format and (n_think_open > 1 or n_think_close > 1):
+        fixes.append(f"多余think标签: {n_think_open}个<think>, {n_think_close}个</think>")
+
+    if not is_new_format:
+        think_matches = re.findall(r"<think>(.*?)</think>", response, re.DOTALL)
+        all_empty = all(m.strip() == "" for m in think_matches) if think_matches else True
+        if all_empty and content:
+            fixes.append("think为空,内容在外部,已合并")
+
+    # ---- 5. 清理 answer ----
+    if gt_answer and answer_block:
+        ans_inner = re.search(r"<answer>(.*?)</answer>", answer_block, re.DOTALL)
+        if ans_inner:
+            ans_text = ans_inner.group(1).strip()
+            if ans_text != gt_answer and len(ans_text) > len(gt_answer):
+                fixes.append(f"answer冗余: \"{ans_text[:50]}\" → \"{gt_answer}\"")
+                answer_block = f"<answer>{gt_answer}</answer>"
+
+    # ---- 6. 重新组装 ----
+    parts = []
+    if content:
+        parts.append(f"<think>\n{content}\n</think>")
+    else:
+        parts.append("<think>\n</think>")
+
+    if answer_block:
+        parts.append(answer_block)
+    if kf_block:
+        parts.append(kf_block)
+
+    cleaned = "\n\n".join(parts)
+    return cleaned, fixes
+
+
+def _extract_option_letter(s: str) -> str:
+    """从答案文本中提取选项字母 A-E，失败返回空串"""
+    m = re.match(r"^\(?([A-E])\)?[\.\s]?", s.strip())
+    return m.group(1) if m else ""
+
+
+_PLACEHOLDERS = {
+    "concise answer", "the answer", "answer", "...", "……",
+    "yes", "no", "n/a", "none", "unknown",
+}
+
+
+def answers_match(model_ans: str, gt: str) -> bool:
+    """答案匹配，分两套策略：
+    - 选择题（双方均可提取 A-E 字母）：严格比较字母
+    - 自由文本：只过滤占位符/空答/极短答案，其余保留
+      （因为 GT 已作为引导传给模型，有实质内容即视为正确）
+    """
+    if not model_ans or not gt:
+        return False
+    if model_ans == gt:
+        return True
+
+    # ---- 选择题：提取字母比较 ----
+    ml = _extract_option_letter(model_ans)
+    gl = _extract_option_letter(gt)
+    if ml and gl:
+        return ml == gl
+
+    # ---- 自由文本：过滤明显失败 ----
+    mn = model_ans.strip().lower()
+    # 占位符
+    if mn in _PLACEHOLDERS:
+        return False
+    # 极短（不足 5 字符）
+    if len(mn) < 5:
+        return False
+    # 纯省略号 / 重复符号
+    if re.match(r"^[.\s…]+$", mn):
+        return False
+
+    # 有实质内容 → 保留（GT 已引导模型，不做严格文本比对）
+    return True
+
+
+def validate_item(item: dict) -> list[str]:
+    """验证单条数据的必要字段"""
+    warnings = []
+    resp = item.get("response", "")
+
+    if not resp:
+        warnings.append("response为空")
+        return warnings
+
+    # 检查 think 内容
+    think_match = re.search(r"<think>(.*?)</think>", resp, re.DOTALL)
+    think_text = think_match.group(1).strip() if think_match else ""
+    if not think_text:
+        warnings.append("think内容为空")
+
+    # 检查 answer
+    if not re.search(r"<answer>", resp):
+        warnings.append("缺少answer")
+
+    # 检查 Reference frame
+    if not re.search(r"Reference frame at", resp):
+        warnings.append("无Reference frame")
+
+    # 检查 delta
+    if not re.search(r"<delta>", resp):
+        warnings.append("无delta")
+
+    # 检查 time 标签
+    n_time = len(re.findall(r"<time>", resp))
+    n_time_range = len(re.findall(r"<time_range>", resp))
+    if n_time + n_time_range == 0:
+        warnings.append("无时间标签")
+
+    # 检查 answer 是否与 GT 一致（宽松匹配）
+    gt = item.get("answer", "")
+    if gt:
+        ans_match = re.search(r"<answer>(.*?)</answer>", resp, re.DOTALL)
+        model_ans = ans_match.group(1).strip() if ans_match else ""
+        if model_ans and not answers_match(model_ans, gt):
+            warnings.append(f"答案错误: model={model_ans[:40]} gt={gt[:40]}")
+
+    return warnings
+
+
+def main():
+    parser = argparse.ArgumentParser(description="清洗 CoT 数据")
+    parser.add_argument("input", help="输入 JSONL 文件路径")
+    parser.add_argument("-o", "--output", help="输出路径 (不指定则覆盖原文件)")
+    parser.add_argument("--filter-wrong", action="store_true",
+                        help="过滤掉答案与GT不一致的条目")
+    args = parser.parse_args()
+
+    input_path = Path(args.input)
+    output_path = Path(args.output) if args.output else input_path
+
+    # 读取
+    items = []
+    with open(input_path, encoding="utf-8") as f:
+        for line in f:
+            if line.strip():
+                items.append(json.loads(line))
+
+    print(f"读取 {len(items)} 条数据\n")
+
+    # 清洗
+    total_fixes = 0
+    total_warnings = 0
+    cleaned_items = []
+
+    for i, item in enumerate(items):
+        pid = item.get("problem_id", i)
+        resp = item.get("response", "")
+
+        if not resp:
+            cleaned_items.append(item)
+            continue
+
+        # 清洗 response
+        gt_answer = item.get("answer", "")
+        cleaned_resp, fixes = clean_response(resp, gt_answer)
+        item["response"] = cleaned_resp
+
+        # 验证
+        warnings = validate_item(item)
+
+        if fixes or warnings:
+            print(f"  [{pid}]", end="")
+            if fixes:
+                total_fixes += len(fixes)
+                print(f"  修复: {'; '.join(fixes)}", end="")
+            if warnings:
+                total_warnings += len(warnings)
+                print(f"  ⚠ {'; '.join(warnings)}", end="")
+            print()
+
+        cleaned_items.append(item)
+
+    # 过滤答案错误的条目
+    n_filtered = 0
+    if args.filter_wrong:
+        filtered = []
+        for item in cleaned_items:
+            gt = item.get("answer", "")
+            resp = item.get("response", "")
+            ans_m = re.search(r"<answer>(.*?)</answer>", resp, re.DOTALL)
+            model_ans = ans_m.group(1).strip() if ans_m else ""
+            if not gt or answers_match(model_ans, gt):
+                filtered.append(item)
+            else:
+                n_filtered += 1
+        cleaned_items = filtered
+
+    # 写入
+    with open(output_path, "w", encoding="utf-8") as f:
+        for item in cleaned_items:
+            f.write(json.dumps(item, ensure_ascii=False) + "\n")
+
+    # 报告
+    print(f"\n{'='*50}")
+    print(f"清洗完成:")
+    print(f"  总条数:   {len(cleaned_items)}")
+    print(f"  修复数:   {total_fixes}")
+    print(f"  警告数:   {total_warnings}")
+    if args.filter_wrong:
+        print(f"  过滤错误: {n_filtered}")
+    print(f"  输出到:   {output_path}")
+
+
+if __name__ == "__main__":
+    main()

cot_generate/generate_video_cot.py

@@ -0,0 +1,567 @@
+"""
+批量视频 CoT 数据生成脚本
+通过 OpenAI 兼容 API 调用 ms-swift 部署的 Qwen3.5-122B-A10B 模型,
+对视频生成带时间戳的 Chain-of-Thought 推理数据。
+
+使用方式:
+  1. 先启动服务: bash start_server.sh
+  2. 输入为视频目录 (无答案, 自由描述模式):
+     python generate_video_cot.py --input /path/to/videos --output output.jsonl
+  3. 输入为 jsonl 文件 (支持 answer 字段引导推理):
+     python generate_video_cot.py --input data.jsonl --output output.jsonl
+  4. 输入为 Video-R1 格式 json 文件:
+     python generate_video_cot.py --input Video-R1-260k.json --output output.jsonl --video_root /path/to/videos
+  5. 支持断点续跑:
+     python generate_video_cot.py --input data.jsonl --output output.jsonl --resume
+
+输入 jsonl 格式 (每行一个 json):
+  {"video": "/path/to/video.mp4", "question": "What is the velocity?", "answer": "3 m/s"}
+  answer 字段可选; 有 answer 时模型会生成指向该答案的推理链
+
+Video-R1 json 格式 (自动转换):
+  [{"path": "...", "problem": "...", "options": [...], "solution": "<answer>A</answer>"}, ...]
+"""
+
+import base64
+import json
+import argparse
+import math
+import os
+import re
+import time
+from pathlib import Path
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+import cv2
+
+from openai import OpenAI
+from tqdm import tqdm
+
+# ============================================================
+# CoT System Prompt — Reference-Frame + Delta (Ref-Δ) Paradigm
+#
+# 核心思想: 用 "参考系(锚点帧) + 动态变化量(Δ)" 的范式来对抗
+# 静态先验捷径(Static Prior Shortcuts)。
+# 模型必须先建立视觉锚点, 再显式追踪相对于锚点的变化(delta),
+# 从而迫使推理基于真实视频演变, 而非背景+语言常识盲猜。
+#
+# 提取方式:
+#   锚点帧: re.findall(r"Reference frame at <time>(.*?)</time>:\s*(.*?)(?=Reference frame at|$)", text, re.DOTALL)
+#   变化量: re.findall(r"<time_range>(.*?)</time_range>.*?<delta>(.*?)</delta>", text, re.DOTALL)
+# ============================================================
+
+# ===================== v4 prompt (已弃用, 保留备份) =====================
+# SYSTEM_PROMPT_GUIDED_V4 = """\
+# Adopt the persona of an expert video analyst tasked with uncovering the true "Semantic Flow" of a video. \
+# Your primary goal is to counteract common AI flaws: "single-frame bias" (jumping to conclusions from \
+# one static image) and "background bias" (over-relying on the scene's setting rather than the actual action). \
+# To do this, you must focus exclusively on the dynamic visual changes (deltas) across time, \
+# strictly ignoring static background elements unless they interact with the main subject. \
+# CRITICAL: The video frames have a precise timestamp watermarked in the top-left corner \
+# in the format of HH:MM:SS. You MUST read these watermarked timestamps directly from \
+# the video frames and use ONLY the exact timestamps you actually observe in the watermark. \
+# Do NOT estimate, interpolate, round, or fabricate any timestamps — every timestamp you \
+# cite in <time> or <time_range> tags must be one you can literally see on a video frame.
+# {fps_constraint}
+# Formatting Rules:
+# 1. Thinking Block: Write your detailed thought process within <think> and </think> tags. \
+# Present a continuous, logical analysis of the video's progression, moving smoothly from \
+# the beginning to the end without abrupt shifts or artificial conversational fillers.
+# 2. Cohesive Prose: Your thought process must be written in flowing paragraphs. \
+# Avoid disjointed bullet points or fragmented lists to ensure a complete, unbroken narrative of the events.
+# 3. Reference Frame (Anchor): When you first observe the scene, or when the scene changes \
+# significantly (cut, new subject, environment shift), establish an anchor using this exact pattern: \
+# "Reference frame at <time>HH:MM:SS</time>: [static description of subject positions, postures, \
+# held objects, spatial relationships at this exact moment]." \
+# The timestamp inside <time> MUST be copied exactly from the watermark visible on that frame — \
+# do not guess or approximate. \
+# The first reference frame should appear at the beginning of your analysis. \
+# You may establish multiple reference frames throughout the video when the scene resets or shifts.
+# 4. Dynamic Change (Delta): When describing how things change over time, use this pattern: \
+# "From <time_range>HH:MM:SS - HH:MM:SS</time_range>, <delta>[description of what changed, \
+# moved, or transformed relative to the most recent reference frame]</delta>." \
+# Both the start and end timestamps in <time_range> MUST be exact watermarked timestamps \
+# you can see on actual video frames — never interpolate between frames or invent timestamps. \
+# Focus only on what changed — do not describe static background. \
+# Each delta should clearly connect back to its preceding reference frame.
+# 5. Anti-Bias Continuity: Build your reasoning seamlessly from one reference frame and its \
+# deltas to the next. Do not describe stationary room decor. Prove your final conclusion by \
+# connecting the sequence of temporal deltas into a complete logical chain.
+# 6. Answer Block: After the </think> block, provide the final concise answer inside \
+# <answer> and </answer> tags.
+# 7. Key Frame/Range Block: After the answer, include a block using <key_frame> and </key_frame> \
+# tags where you point out the specific timestamp(s) or time range(s) of the critical video frames \
+# without which the question cannot be answered.
+#
+# Example output structure:
+#
+# <think>
+# Reference frame at <time>00:00:00</time>: a person stands at a table holding a red cup \
+# in their right hand, with a white bowl positioned to the left.
+#
+# The action begins immediately. From <time_range>00:00:00 - 00:00:02</time_range>, \
+# <delta>the person lifts the red cup from the table and tilts it toward the bowl on the left</delta>. \
+# The motion is deliberate, suggesting an intentional pouring action.
+#
+# Continuing the flow, from <time_range>00:00:02 - 00:00:04</time_range>, \
+# <delta>the cup is now fully inverted and liquid streams steadily into the bowl</delta>. \
+# This confirms the pouring interpretation.
+#
+# The scene then resets. Reference frame at <time>00:00:05</time>: the cup is back upright \
+# on the table, and the bowl is now visibly full of liquid.
+#
+# From <time_range>00:00:05 - 00:00:07</time_range>, \
+# <delta>the person pushes the bowl forward across the table toward the opposite edge</delta>.
+#
+# Connecting the delta chain: the person poured liquid from cup to bowl, then pushed the bowl away.
+# </think>
+#
+# <answer>The person poured liquid from the cup into the bowl.</answer>
+#
+# <key_frame>00:00:02 - 00:00:04, 00:00:05</key_frame>
+#
+# Here is your question:
+# {question}
+#
+# The final answer is provided below. Use it as a reference to guide your thinking process. \
+# You MUST logically arrive at this correct conclusion by connecting the dynamic temporal \
+# deltas in a continuous, well-reasoned flow.
+# {answer}"""
+# ===================== end v4 prompt =====================
+
+# 有正确答案时使用 (answer-guided) — v5 精简版
+SYSTEM_PROMPT_GUIDED = """\
+You are an expert video analyst. Analyze the "Semantic Flow" of a video by \
+tracking dynamic changes across time. Ignore static backgrounds unless they \
+interact with the main subject.
+
+CRITICAL: Video frames have timestamps watermarked in the top-left corner. \
+Read timestamps directly from the watermark — never estimate or fabricate. \
+Always round DOWN to whole seconds and write timestamps as HH:MM:SS only \
+(e.g. if watermark shows 00:00:04.8, write 00:00:04).
+{fps_constraint}
+Output format:
+1. Reference Frame: "Reference frame at <time>HH:MM:SS</time>: [brief description]."
+2. Delta: "From <time_range>HH:MM:SS - HH:MM:SS</time_range>, <delta>[what changed]</delta>."
+3. Reasoning Summary: "<reasoning_summary>[connect deltas to answer]</reasoning_summary>"
+4. Answer: "<answer>concise answer</answer>"
+
+Example:
+Reference frame at <time>00:00:00</time>: a person holds a red cup, white bowl to the left.
+From <time_range>00:00:00 - 00:00:03</time_range>, <delta>the person tilts the cup and pours liquid into the bowl</delta>.
+From <time_range>00:00:03 - 00:00:05</time_range>, <delta>the person pushes the filled bowl forward</delta>.
+<reasoning_summary>The person poured liquid from cup to bowl, then pushed it away.</reasoning_summary>
+<answer>The person poured liquid into the bowl.</answer>
+
+Cover the entire video timeline. Each delta must describe a different change — \
+never repeat previous observations. If an action repeats, summarize it in one \
+delta instead of describing each repetition. Keep your analysis concise.
+
+Question: {question}
+
+The correct answer is given below. You MUST logically arrive at this conclusion \
+by connecting the temporal deltas.
+{answer}"""
+
+# 无正确答案时使用 (free reasoning) — v5 精简版
+SYSTEM_PROMPT_FREE = """\
+You are an expert video analyst. Analyze the "Semantic Flow" of a video by \
+tracking dynamic changes across time. Ignore static backgrounds unless they \
+interact with the main subject.
+
+CRITICAL: Video frames have timestamps watermarked in the top-left corner. \
+Read timestamps directly from the watermark — never estimate or fabricate. \
+Always round DOWN to whole seconds and write timestamps as HH:MM:SS only \
+(e.g. if watermark shows 00:00:04.8, write 00:00:04).
+{fps_constraint}
+Output format:
+1. Reference Frame: "Reference frame at <time>HH:MM:SS</time>: [brief description]."
+2. Delta: "From <time_range>HH:MM:SS - HH:MM:SS</time_range>, <delta>[what changed]</delta>."
+3. Reasoning Summary: "<reasoning_summary>[connect deltas to answer]</reasoning_summary>"
+4. Answer: "<answer>concise answer</answer>"
+
+Example:
+Reference frame at <time>00:00:00</time>: a person holds a red cup, white bowl to the left.
+From <time_range>00:00:00 - 00:00:03</time_range>, <delta>the person tilts the cup and pours liquid into the bowl</delta>.
+From <time_range>00:00:03 - 00:00:05</time_range>, <delta>the person pushes the filled bowl forward</delta>.
+<reasoning_summary>The person poured liquid from cup to bowl, then pushed it away.</reasoning_summary>
+<answer>The person poured liquid into the bowl.</answer>
+
+Cover the entire video timeline. Each delta must describe a different change — \
+never repeat previous observations. If an action repeats, summarize it in one \
+delta instead of describing each repetition. Keep your analysis concise.
+
+Question: {question}"""
+
+DEFAULT_QUESTION = "Describe and analyze what is happening in this video step by step."
+
+VIDEO_EXTENSIONS = {".mp4", ".avi", ".mkv", ".mov", ".webm", ".flv"}
+
+MIME_MAP = {
+    ".mp4": "video/mp4", ".avi": "video/avi", ".mkv": "video/x-matroska",
+    ".mov": "video/quicktime", ".webm": "video/webm", ".flv": "video/x-flv",
+}
+
+
+def _smart_nframes(total_frames: int, video_fps: float, max_frames: int,
+                   sample_fps: float = 2.0) -> int:
+    """复现 qwen_vl_utils.vision_process.smart_nframes 的完整逻辑
+
+    ms-swift (Qwen2.5-VL / Qwen3.5) 的实际抽帧流程:
+      1. 按目标采样率(默认2FPS)算出理想帧数: total_frames / video_fps * sample_fps
+      2. clamp 到 [min_frames=4, max_frames] 区间
+      3. 向下取到 FRAME_FACTOR=2 的倍数 (帧数必须是偶数)
+
+    参数:
+      total_frames: 视频原始总帧数
+      video_fps:    视频原始帧率
+      max_frames:   最大采样帧数 (对应 FPS_MAX_FRAMES 环境变量, 默认32)
+      sample_fps:   目标采样帧率 (qwen_vl_utils 默认 2.0)
+    """
+    FRAME_FACTOR = 2
+    min_frames = math.ceil(4 / FRAME_FACTOR) * FRAME_FACTOR     # = 4
+    max_frames = math.floor(min(max_frames, total_frames) / FRAME_FACTOR) * FRAME_FACTOR
+
+    nframes = total_frames / video_fps * sample_fps
+    nframes = min(min(max(nframes, min_frames), max_frames), total_frames)
+    nframes = math.floor(nframes / FRAME_FACTOR) * FRAME_FACTOR
+
+    # 安全边界
+    nframes = max(nframes, FRAME_FACTOR)
+    nframes = min(nframes, total_frames)
+
+    return int(nframes)
+
+
+def _get_video_actual_fps(video_path: str, max_frames: int = 32) -> float:
+    """根据视频实际时长, 用 smart_nframes 计算 ms-swift 的真实采样率"""
+    cap = cv2.VideoCapture(video_path)
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    native_fps = cap.get(cv2.CAP_PROP_FPS)
+    cap.release()
+
+    if native_fps <= 0 or total_frames <= 0:
+        return 2.0  # fallback
+
+    duration = total_frames / native_fps
+    sampled = _smart_nframes(total_frames, native_fps, max_frames)
+    return sampled / duration if duration > 0 else 2.0
+
+
+def _build_fps_constraint(fps: float) -> str:
+    """根据实际抽帧率生成时间戳约束段落 (宽松措辞, 以水印为准)"""
+    interval = 1.0 / fps
+    return (
+        f"\nIMPORTANT: The video frames you see have been sampled at approximately "
+        f"{fps:.1f} FPS (roughly one frame every {interval:.1f}s). "
+        f"The timestamps watermarked on each frame reflect the actual frame times. "
+        f"Always use the exact watermark timestamps you observe — "
+        f"do not invent, interpolate, or round timestamps to values "
+        f"that do not appear in the watermark.\n"
+    )
+
+
+def build_messages(video_path: str, question: str, answer: str = None, max_frames: int = 32) -> list:
+    """构建包含视频的多模态消息 (base64 编码本地文件)
+
+    answer: 若提供, 使用 answer-guided prompt 引导模型推理到正确答案;
+            否则使用 free reasoning prompt 自由推理。
+    question 和 answer 嵌入 system prompt, user 消息仅含视频。
+    max_frames: 对应 FPS_MAX_FRAMES, 用于动态计算实际采样率。
+    """
+    path = Path(video_path).resolve()
+    mime = MIME_MAP.get(path.suffix.lower(), "video/mp4")
+    with open(path, "rb") as f:
+        b64 = base64.b64encode(f.read()).decode("utf-8")
+    video_url = f"data:{mime};base64,{b64}"
+
+    actual_fps = _get_video_actual_fps(video_path, max_frames)
+    fps_constraint = _build_fps_constraint(actual_fps)
+
+    if answer is not None:
+        system = SYSTEM_PROMPT_GUIDED.format(question=question, answer=answer, fps_constraint=fps_constraint)
+    else:
+        system = SYSTEM_PROMPT_FREE.format(question=question, fps_constraint=fps_constraint)
+
+    return [
+        {"role": "system", "content": system},
+        {
+            "role": "user",
+            "content": [
+                {"type": "video_url", "video_url": {"url": video_url}},
+                {"type": "text", "text": "Please analyze the video and answer the question."},
+            ],
+        },
+    ]
+
+
+def generate_single(
+    client: OpenAI,
+    model: str,
+    video_path: str,
+    question: str,
+    max_tokens: int,
+    temperature: float,
+    answer: str = None,
+    max_frames: int = 32,
+) -> str:
+    """对单个视频生成 CoT"""
+    messages = build_messages(video_path, question, answer, max_frames=max_frames)
+    response = client.chat.completions.create(
+        model=model,
+        messages=messages,
+        max_tokens=max_tokens,
+        temperature=temperature,
+    )
+    return response.choices[0].message.content
+
+
+# ============================================================
+# Video-R1 格式转换
+# ============================================================
+def _extract_answer(solution: str) -> str:
+    """从 '<answer>A</answer>' 中提取答案文本"""
+    m = re.search(r"<answer>(.*?)</answer>", solution, re.DOTALL)
+    return m.group(1).strip() if m else solution.strip()
+
+
+def load_video_r1(json_path: str, video_root: str, default_question: str) -> list:
+    """加载 Video-R1 格式 json, 转换为统一 item 格式"""
+    root = Path(video_root) if video_root else Path(".")
+    with open(json_path, encoding="utf-8") as f:
+        data = json.load(f)
+    items = []
+    for entry in data:
+        # 视频路径: video_root / entry["path"] (去掉开头的 ./)
+        rel = entry.get("path", "").lstrip("./")
+        video = str(root / rel)
+
+        # 问题: problem + options
+        problem = entry.get("problem", default_question)
+        options = entry.get("options", [])
+        if options:
+            question = problem.rstrip() + "\n" + "\n".join(options)
+        else:
+            question = problem
+
+        # 答案
+        solution = entry.get("solution", "")
+        answer = _extract_answer(solution) if solution else None
+
+        item = {"video": video, "question": question}
+        if answer:
+            item["answer"] = answer
+        # 保留原始字段供输出追溯
+        if "problem_id" in entry:
+            item["problem_id"] = entry["problem_id"]
+        items.append(item)
+    return items
+
+
+# ============================================================
+# 通用输入加载
+# ============================================================
+def load_inputs(input_path: str, default_question: str, video_root: str = None) -> list:
+    """
+    支持格式:
+    1. 视频文件目录
+    2. 单个视频文件
+    3. jsonl 文件: 每行 {"video", "question"[, "answer"]}
+    4. Video-R1 json 文件: [{"path", "problem", "options", "solution"}, ...]
+    """
+    path = Path(input_path)
+    items = []
+
+    if path.is_dir():
+        for f in sorted(path.rglob("*")):
+            if f.suffix.lower() in VIDEO_EXTENSIONS:
+                items.append({"video": str(f), "question": default_question})
+
+    elif path.is_file() and path.suffix.lower() in VIDEO_EXTENSIONS:
+        items.append({"video": str(path), "question": default_question})
+
+    elif path.suffix == ".jsonl":
+        with open(path, encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                item = json.loads(line)
+                item.setdefault("question", default_question)
+                items.append(item)
+
+    elif path.suffix == ".json":
+        with open(path, encoding="utf-8") as f:
+            data = json.load(f)
+        if not isinstance(data, list):
+            raise ValueError(f"JSON file should contain a list, got {type(data)}")
+        # 自动检测是否为 Video-R1 格式 (含 "path" 和 "problem" 字段)
+        if data and "path" in data[0] and "problem" in data[0]:
+            items = load_video_r1(input_path, video_root or ".", default_question)
+        else:
+            for item in data:
+                item.setdefault("question", default_question)
+                items.append(item)
+    else:
+        raise ValueError(f"Unsupported input format: {path.suffix}")
+
+    return items
+
+
+def load_processed(output_path: str) -> set:
+    """加载已处理的条目 key (用于断点续跑)
+    key = video + question，支持同一视频多个不同问题"""
+    processed = set()
+    if Path(output_path).exists():
+        with open(output_path, encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                entry = json.loads(line)
+                video    = entry.get("video", "")
+                question = entry.get("question", "")
+                processed.add((video, question))
+    return processed
+
+
+# ============================================================
+# 处理流程
+# ============================================================
+def _make_result(item: dict) -> dict:
+    result = {"video": item["video"], "question": item["question"]}
+    if "answer" in item:
+        result["answer"] = item["answer"]
+    if "problem_id" in item:
+        result["problem_id"] = item["problem_id"]
+    if "data_source" in item:
+        result["data_source"] = item["data_source"]
+    return result
+
+
+def process_sequential(client, model, items, output_path, max_tokens, temperature, mode, max_frames):
+    with open(output_path, mode, encoding="utf-8") as fout:
+        for item in tqdm(items, desc="Generating CoT"):
+            result = _make_result(item)
+            try:
+                result["response"] = generate_single(
+                    client, model,
+                    item["video"], item["question"],
+                    max_tokens, temperature,
+                    answer=item.get("answer"),
+                    max_frames=max_frames,
+                )
+            except Exception as e:
+                print(f"\n[ERROR] {item['video']}: {e}")
+                result["error"] = str(e)
+            fout.write(json.dumps(result, ensure_ascii=False) + "\n")
+            fout.flush()
+
+
+def process_parallel(client, model, items, output_path, max_tokens, temperature, mode, workers, max_frames):
+    with open(output_path, mode, encoding="utf-8") as fout:
+        with ThreadPoolExecutor(max_workers=workers) as executor:
+            futures = {
+                executor.submit(
+                    generate_single,
+                    client, model,
+                    item["video"], item["question"],
+                    max_tokens, temperature,
+                    item.get("answer"),
+                    max_frames,
+                ): item
+                for item in items
+            }
+            for future in tqdm(as_completed(futures), total=len(futures), desc="Generating CoT"):
+                item = futures[future]
+                result = _make_result(item)
+                try:
+                    result["response"] = future.result()
+                except Exception as e:
+                    print(f"\n[ERROR] {item['video']}: {e}")
+                    result["error"] = str(e)
+                fout.write(json.dumps(result, ensure_ascii=False) + "\n")
+                fout.flush()
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="批量视频 CoT 数据生成 (Qwen3.5 + ms-swift)",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument("--input", required=True,
+                        help="输入: 视频目录 / jsonl 文件 / json 文件 (含 Video-R1 格式)")
+    parser.add_argument("--output", required=True,
+                        help="输出 jsonl 文件路径")
+    parser.add_argument("--video_root", default=None,
+                        help="Video-R1 json 格式时, 视频文件的根目录")
+    parser.add_argument("--api_base", default="http://localhost:8000/v1",
+                        help="ms-swift deploy 的 API 地址 (default: http://localhost:8000/v1)")
+    parser.add_argument("--model", default="Qwen3.5-122B-A10B",
+                        help="模型名称, 需与 swift deploy 一致")
+    parser.add_argument("--question", default=DEFAULT_QUESTION,
+                        help="默认问题 (当输入数据无 question 字段时使用)")
+    parser.add_argument("--max_tokens", type=int, default=4096,
+                        help="最大生成 token 数 (default: 4096)")
+    parser.add_argument("--temperature", type=float, default=0.7,
+                        help="生成温度 (default: 0.7)")
+    parser.add_argument("--workers", type=int, default=1,
+                        help="并行请求数 (default: 1, vLLM 会自动做 batch)")
+    parser.add_argument("--max_frames", type=int, default=None,
+                        help="最大采样帧数, 需与 FPS_MAX_FRAMES 环境变量一致 (默认读取环境变量, 否则 32)")
+    parser.add_argument("--resume", action="store_true",
+                        help="断点续跑: 跳过已处理的视频")
+    args = parser.parse_args()
+
+    # max_frames: 优先用 CLI 参数, 其次读环境变量 FPS_MAX_FRAMES, 最后默认 32
+    max_frames = args.max_frames or int(os.environ.get("FPS_MAX_FRAMES", 32))
+
+    items = load_inputs(args.input, args.question, args.video_root)
+    print(f"[INFO] Loaded {len(items)} items")
+    has_answer = sum(1 for i in items if "answer" in i)
+    print(f"[INFO] Items with ground-truth answer: {has_answer}/{len(items)}")
+    print(f"[INFO] Max sampled frames per video: {max_frames} (actual FPS varies by video duration)")
+
+    mode = "w"
+    if args.resume:
+        processed = load_processed(args.output)
+        before = len(items)
+        items = [item for item in items if (item["video"], item["question"]) not in processed]
+        print(f"[INFO] Resume mode: skipped {before - len(items)}, remaining {len(items)}")
+        mode = "a"
+
+    if not items:
+        print("[INFO] Nothing to process. Done!")
+        return
+
+    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
+
+    client = OpenAI(api_key="EMPTY", base_url=args.api_base)
+    print(f"[INFO] Connecting to {args.api_base} ...")
+    try:
+        models = client.models.list()
+        print(f"[INFO] Server ready. Available models: {[m.id for m in models.data]}")
+    except Exception as e:
+        print(f"[ERROR] Cannot connect to server: {e}")
+        print("[HINT] Please run 'bash start_server.sh' first.")
+        return
+
+    t0 = time.time()
+    if args.workers <= 1:
+        process_sequential(client, args.model, items, args.output,
+                           args.max_tokens, args.temperature, mode, max_frames)
+    else:
+        process_parallel(client, args.model, items, args.output,
+                         args.max_tokens, args.temperature, mode, args.workers, max_frames)
+
+    elapsed = time.time() - t0
+    print(f"\n[DONE] Results saved to {args.output}")
+    print(f"[DONE] Processed {len(items)} videos in {elapsed:.1f}s "
+          f"({elapsed / len(items):.1f}s per video)")
+
+
+if __name__ == "__main__":
+    main()
+
+
+# tmux send-keys -t cot_gen C-c 2>/dev/null; sleep 1; tmux send-keys -t cot_gen "cd /root/zwj/CoT_generate && python generate_video_cot.py --input /root/zwj/CoT_generate/output/need_rerun_v2.jsonl --output /root/zwj/CoT_generate/output/need_rerun_v2_cot.jsonl --max_frames 768 --workers 4 --resume" Enter

cot_generate/generate_video_cot_with_hint.py

@@ -0,0 +1,581 @@
+"""
+批量视频 CoT 数据生成脚本
+通过 OpenAI 兼容 API 调用 ms-swift 部署的 Qwen3.5-122B-A10B 模型,
+对视频生成带时间戳的 Chain-of-Thought 推理数据。
+
+使用方式:
+  1. 先启动服务: bash start_server.sh
+  2. 输入为视频目录 (无答案, 自由描述模式):
+     python generate_video_cot.py --input /path/to/videos --output output.jsonl
+  3. 输入为 jsonl 文件 (支持 answer 字段引导推理):
+     python generate_video_cot.py --input data.jsonl --output output.jsonl
+  4. 输入为 Video-R1 格式 json 文件:
+     python generate_video_cot.py --input Video-R1-260k.json --output output.jsonl --video_root /path/to/videos
+  5. 支持断点续跑:
+     python generate_video_cot.py --input data.jsonl --output output.jsonl --resume
+
+输入 jsonl 格式 (每行一个 json):
+  {"video": "/path/to/video.mp4", "question": "What is the velocity?", "answer": "3 m/s"}
+  answer 字段可选; 有 answer 时模型会生成指向该答案的推理链
+
+Video-R1 json 格式 (自动转换):
+  [{"path": "...", "problem": "...", "options": [...], "solution": "<answer>A</answer>"}, ...]
+"""
+
+import base64
+import json
+import argparse
+import math
+import os
+import re
+import time
+from pathlib import Path
+from concurrent.futures import ThreadPoolExecutor, as_completed
+
+import cv2
+
+from openai import OpenAI
+from tqdm import tqdm
+
+# ============================================================
+# CoT System Prompt — Reference-Frame + Delta (Ref-Δ) Paradigm
+#
+# 核心思想: 用 "参考系(锚点帧) + 动态变化量(Δ)" 的范式来对抗
+# 静态先验捷径(Static Prior Shortcuts)。
+# 模型必须先建立视觉锚点, 再显式追踪相对于锚点的变化(delta),
+# 从而迫使推理基于真实视频演变, 而非背景+语言常识盲猜。
+#
+# 提取方式:
+#   锚点帧: re.findall(r"Reference frame at <time>(.*?)</time>:\s*(.*?)(?=Reference frame at|$)", text, re.DOTALL)
+#   变化量: re.findall(r"<time_range>(.*?)</time_range>.*?<delta>(.*?)</delta>", text, re.DOTALL)
+# ============================================================
+
+# ===================== v4 prompt (已弃用, 保留备份) =====================
+# SYSTEM_PROMPT_GUIDED_V4 = """\
+# Adopt the persona of an expert video analyst tasked with uncovering the true "Semantic Flow" of a video. \
+# Your primary goal is to counteract common AI flaws: "single-frame bias" (jumping to conclusions from \
+# one static image) and "background bias" (over-relying on the scene's setting rather than the actual action). \
+# To do this, you must focus exclusively on the dynamic visual changes (deltas) across time, \
+# strictly ignoring static background elements unless they interact with the main subject. \
+# CRITICAL: The video frames have a precise timestamp watermarked in the top-left corner \
+# in the format of HH:MM:SS. You MUST read these watermarked timestamps directly from \
+# the video frames and use ONLY the exact timestamps you actually observe in the watermark. \
+# Do NOT estimate, interpolate, round, or fabricate any timestamps — every timestamp you \
+# cite in <time> or <time_range> tags must be one you can literally see on a video frame.
+# {fps_constraint}
+# Formatting Rules:
+# 1. Thinking Block: Write your detailed thought process within <think> and </think> tags. \
+# Present a continuous, logical analysis of the video's progression, moving smoothly from \
+# the beginning to the end without abrupt shifts or artificial conversational fillers.
+# 2. Cohesive Prose: Your thought process must be written in flowing paragraphs. \
+# Avoid disjointed bullet points or fragmented lists to ensure a complete, unbroken narrative of the events.
+# 3. Reference Frame (Anchor): When you first observe the scene, or when the scene changes \
+# significantly (cut, new subject, environment shift), establish an anchor using this exact pattern: \
+# "Reference frame at <time>HH:MM:SS</time>: [static description of subject positions, postures, \
+# held objects, spatial relationships at this exact moment]." \
+# The timestamp inside <time> MUST be copied exactly from the watermark visible on that frame — \
+# do not guess or approximate. \
+# The first reference frame should appear at the beginning of your analysis. \
+# You may establish multiple reference frames throughout the video when the scene resets or shifts.
+# 4. Dynamic Change (Delta): When describing how things change over time, use this pattern: \
+# "From <time_range>HH:MM:SS - HH:MM:SS</time_range>, <delta>[description of what changed, \
+# moved, or transformed relative to the most recent reference frame]</delta>." \
+# Both the start and end timestamps in <time_range> MUST be exact watermarked timestamps \
+# you can see on actual video frames — never interpolate between frames or invent timestamps. \
+# Focus only on what changed — do not describe static background. \
+# Each delta should clearly connect back to its preceding reference frame.
+# 5. Anti-Bias Continuity: Build your reasoning seamlessly from one reference frame and its \
+# deltas to the next. Do not describe stationary room decor. Prove your final conclusion by \
+# connecting the sequence of temporal deltas into a complete logical chain.
+# 6. Answer Block: After the </think> block, provide the final concise answer inside \
+# <answer> and </answer> tags.
+# 7. Key Frame/Range Block: After the answer, include a block using <key_frame> and </key_frame> \
+# tags where you point out the specific timestamp(s) or time range(s) of the critical video frames \
+# without which the question cannot be answered.
+#
+# Example output structure:
+#
+# <think>
+# Reference frame at <time>00:00:00</time>: a person stands at a table holding a red cup \
+# in their right hand, with a white bowl positioned to the left.
+#
+# The action begins immediately. From <time_range>00:00:00 - 00:00:02</time_range>, \
+# <delta>the person lifts the red cup from the table and tilts it toward the bowl on the left</delta>. \
+# The motion is deliberate, suggesting an intentional pouring action.
+#
+# Continuing the flow, from <time_range>00:00:02 - 00:00:04</time_range>, \
+# <delta>the cup is now fully inverted and liquid streams steadily into the bowl</delta>. \
+# This confirms the pouring interpretation.
+#
+# The scene then resets. Reference frame at <time>00:00:05</time>: the cup is back upright \
+# on the table, and the bowl is now visibly full of liquid.
+#
+# From <time_range>00:00:05 - 00:00:07</time_range>, \
+# <delta>the person pushes the bowl forward across the table toward the opposite edge</delta>.
+#
+# Connecting the delta chain: the person poured liquid from cup to bowl, then pushed the bowl away.
+# </think>
+#
+# <answer>The person poured liquid from the cup into the bowl.</answer>
+#
+# <key_frame>00:00:02 - 00:00:04, 00:00:05</key_frame>
+#
+# Here is your question:
+# {question}
+#
+# The final answer is provided below. Use it as a reference to guide your thinking process. \
+# You MUST logically arrive at this correct conclusion by connecting the dynamic temporal \
+# deltas in a continuous, well-reasoned flow.
+# {answer}"""
+# ===================== end v4 prompt =====================
+
+# 有正确答案时使用 (answer-guided) — v5 精简版
+SYSTEM_PROMPT_GUIDED = """\
+You are an expert video analyst. Analyze the "Semantic Flow" of a video by \
+tracking dynamic changes across time. Ignore static backgrounds unless they \
+interact with the main subject.
+
+CRITICAL: Video frames have HH:MM:SS timestamps watermarked in the top-left \
+corner. Use ONLY exact timestamps you observe — never estimate or fabricate.
+{fps_constraint}
+Output format:
+1. Reference Frame: "Reference frame at <time>HH:MM:SS</time>: [brief description]."
+2. Delta: "From <time_range>HH:MM:SS - HH:MM:SS</time_range>, <delta>[what changed]</delta>."
+3. Reasoning Summary: "<reasoning_summary>[connect deltas to answer]</reasoning_summary>"
+4. Answer: "<answer>concise answer</answer>"
+
+Example:
+Reference frame at <time>00:00:00</time>: a person holds a red cup, white bowl to the left.
+From <time_range>00:00:00 - 00:00:03</time_range>, <delta>the person tilts the cup and pours liquid into the bowl</delta>.
+From <time_range>00:00:03 - 00:00:05</time_range>, <delta>the person pushes the filled bowl forward</delta>.
+<reasoning_summary>The person poured liquid from cup to bowl, then pushed it away.</reasoning_summary>
+<answer>The person poured liquid into the bowl.</answer>
+
+Cover the entire video timeline. Each delta must describe a different change — \
+never repeat previous observations. If an action repeats, summarize it in one \
+delta instead of describing each repetition. Keep your analysis concise.
+
+{hint}Question: {question}
+
+The correct answer is given below. You MUST logically arrive at this conclusion \
+by connecting the temporal deltas.
+{answer}"""
+
+# 无正确答案时使用 (free reasoning) — v5 精简版
+SYSTEM_PROMPT_FREE = """\
+You are an expert video analyst. Analyze the "Semantic Flow" of a video by \
+tracking dynamic changes across time. Ignore static backgrounds unless they \
+interact with the main subject.
+
+CRITICAL: Video frames have HH:MM:SS timestamps watermarked in the top-left \
+corner. Use ONLY exact timestamps you observe — never estimate or fabricate.
+{fps_constraint}
+Output format:
+1. Reference Frame: "Reference frame at <time>HH:MM:SS</time>: [brief description]."
+2. Delta: "From <time_range>HH:MM:SS - HH:MM:SS</time_range>, <delta>[what changed]</delta>."
+3. Reasoning Summary: "<reasoning_summary>[connect deltas to answer]</reasoning_summary>"
+4. Answer: "<answer>concise answer</answer>"
+
+Example:
+Reference frame at <time>00:00:00</time>: a person holds a red cup, white bowl to the left.
+From <time_range>00:00:00 - 00:00:03</time_range>, <delta>the person tilts the cup and pours liquid into the bowl</delta>.
+From <time_range>00:00:03 - 00:00:05</time_range>, <delta>the person pushes the filled bowl forward</delta>.
+<reasoning_summary>The person poured liquid from cup to bowl, then pushed it away.</reasoning_summary>
+<answer>The person poured liquid into the bowl.</answer>
+
+Cover the entire video timeline. Each delta must describe a different change — \
+never repeat previous observations. If an action repeats, summarize it in one \
+delta instead of describing each repetition. Keep your analysis concise.
+
+{hint}Question: {question}"""
+
+# ============================================================
+# Dimension-specific hints (inserted before Question:)
+# ============================================================
+HINT_SPEED     = "NOTE: This is a speed-related question. Your analysis should reflect how fast or slow the action is happening in the video.\n"
+HINT_DIRECTION = "NOTE: This is a direction-related question. Your analysis should reflect which direction the subject is moving in the video.\n"
+HINT_TEMPORAL  = "NOTE: This question asks about temporal order. Your analysis should reflect whether the sequence of events flows naturally forward or appears reversed.\n"
+
+DEFAULT_QUESTION = "Describe and analyze what is happening in this video step by step."
+
+VIDEO_EXTENSIONS = {".mp4", ".avi", ".mkv", ".mov", ".webm", ".flv"}
+
+MIME_MAP = {
+    ".mp4": "video/mp4", ".avi": "video/avi", ".mkv": "video/x-matroska",
+    ".mov": "video/quicktime", ".webm": "video/webm", ".flv": "video/x-flv",
+}
+
+
+def _smart_nframes(total_frames: int, video_fps: float, max_frames: int,
+                   sample_fps: float = 2.0) -> int:
+    """复现 qwen_vl_utils.vision_process.smart_nframes 的完整逻辑
+
+    ms-swift (Qwen2.5-VL / Qwen3.5) 的实际抽帧流程:
+      1. 按目标采样率(默认2FPS)算出理想帧数: total_frames / video_fps * sample_fps
+      2. clamp 到 [min_frames=4, max_frames] 区间
+      3. 向下取到 FRAME_FACTOR=2 的倍数 (帧数必须是偶数)
+
+    参数:
+      total_frames: 视频原始总帧数
+      video_fps:    视频原始帧率
+      max_frames:   最大采样帧数 (对应 FPS_MAX_FRAMES 环境变量, 默认32)
+      sample_fps:   目标采样帧率 (qwen_vl_utils 默认 2.0)
+    """
+    FRAME_FACTOR = 2
+    min_frames = math.ceil(4 / FRAME_FACTOR) * FRAME_FACTOR     # = 4
+    max_frames = math.floor(min(max_frames, total_frames) / FRAME_FACTOR) * FRAME_FACTOR
+
+    nframes = total_frames / video_fps * sample_fps
+    nframes = min(min(max(nframes, min_frames), max_frames), total_frames)
+    nframes = math.floor(nframes / FRAME_FACTOR) * FRAME_FACTOR
+
+    # 安全边界
+    nframes = max(nframes, FRAME_FACTOR)
+    nframes = min(nframes, total_frames)
+
+    return int(nframes)
+
+
+def _get_video_actual_fps(video_path: str, max_frames: int = 32) -> float:
+    """根据视频实际时长, 用 smart_nframes 计算 ms-swift 的真实采样率"""
+    cap = cv2.VideoCapture(video_path)
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    native_fps = cap.get(cv2.CAP_PROP_FPS)
+    cap.release()
+
+    if native_fps <= 0 or total_frames <= 0:
+        return 2.0  # fallback
+
+    duration = total_frames / native_fps
+    sampled = _smart_nframes(total_frames, native_fps, max_frames)
+    return sampled / duration if duration > 0 else 2.0
+
+
+def _build_fps_constraint(fps: float) -> str:
+    """根据实际抽帧率生成时间戳约束段落 (宽松措辞, 以水印为准)"""
+    interval = 1.0 / fps
+    return (
+        f"\nIMPORTANT: The video frames you see have been sampled at approximately "
+        f"{fps:.1f} FPS (roughly one frame every {interval:.1f}s). "
+        f"The timestamps watermarked on each frame reflect the actual frame times. "
+        f"Always use the exact watermark timestamps you observe — "
+        f"do not invent, interpolate, or round timestamps to values "
+        f"that do not appear in the watermark.\n"
+    )
+
+
+def build_messages(video_path: str, question: str, answer: str = None, max_frames: int = 32, hint: str = "") -> list:
+    """构建包含视频的多模态消息 (base64 编码本地文件)
+
+    answer: 若提供, 使用 answer-guided prompt 引导模型推理到正确答案;
+            否则使用 free reasoning prompt 自由推理。
+    question 和 answer 嵌入 system prompt, user 消息仅含视频。
+    max_frames: 对应 FPS_MAX_FRAMES, 用于动态计算实际采样率。
+    hint: 维度专项提示，插入到 Question: 前面，空字符串表示不加提示。
+    """
+    path = Path(video_path).resolve()
+    mime = MIME_MAP.get(path.suffix.lower(), "video/mp4")
+    with open(path, "rb") as f:
+        b64 = base64.b64encode(f.read()).decode("utf-8")
+    video_url = f"data:{mime};base64,{b64}"
+
+    actual_fps = _get_video_actual_fps(video_path, max_frames)
+    fps_constraint = _build_fps_constraint(actual_fps)
+    hint_str = hint + "\n" if hint else ""
+
+    if answer is not None:
+        system = SYSTEM_PROMPT_GUIDED.format(question=question, answer=answer, fps_constraint=fps_constraint, hint=hint_str)
+    else:
+        system = SYSTEM_PROMPT_FREE.format(question=question, fps_constraint=fps_constraint, hint=hint_str)
+
+    return [
+        {"role": "system", "content": system},
+        {
+            "role": "user",
+            "content": [
+                {"type": "video_url", "video_url": {"url": video_url}},
+                {"type": "text", "text": "Please analyze the video and answer the question."},
+            ],
+        },
+    ]
+
+
+def generate_single(
+    client: OpenAI,
+    model: str,
+    video_path: str,
+    question: str,
+    max_tokens: int,
+    temperature: float,
+    answer: str = None,
+    max_frames: int = 32,
+    hint: str = "",
+) -> str:
+    """对单个视频生成 CoT"""
+    messages = build_messages(video_path, question, answer, max_frames=max_frames, hint=hint)
+    response = client.chat.completions.create(
+        model=model,
+        messages=messages,
+        max_tokens=max_tokens,
+        temperature=temperature,
+    )
+    return response.choices[0].message.content
+
+
+# ============================================================
+# Video-R1 格式转换
+# ============================================================
+def _extract_answer(solution: str) -> str:
+    """从 '<answer>A</answer>' 中提取答案文本"""
+    m = re.search(r"<answer>(.*?)</answer>", solution, re.DOTALL)
+    return m.group(1).strip() if m else solution.strip()
+
+
+def load_video_r1(json_path: str, video_root: str, default_question: str) -> list:
+    """加载 Video-R1 格式 json, 转换为统一 item 格式"""
+    root = Path(video_root) if video_root else Path(".")
+    with open(json_path, encoding="utf-8") as f:
+        data = json.load(f)
+    items = []
+    for entry in data:
+        # 视频路径: video_root / entry["path"] (去掉开头的 ./)
+        rel = entry.get("path", "").lstrip("./")
+        video = str(root / rel)
+
+        # 问题: problem + options
+        problem = entry.get("problem", default_question)
+        options = entry.get("options", [])
+        if options:
+            question = problem.rstrip() + "\n" + "\n".join(options)
+        else:
+            question = problem
+
+        # 答案
+        solution = entry.get("solution", "")
+        answer = _extract_answer(solution) if solution else None
+
+        item = {"video": video, "question": question}
+        if answer:
+            item["answer"] = answer
+        # 保留原始字段供输出追溯
+        if "problem_id" in entry:
+            item["problem_id"] = entry["problem_id"]
+        items.append(item)
+    return items
+
+
+# ============================================================
+# 通用输入加载
+# ============================================================
+def load_inputs(input_path: str, default_question: str, video_root: str = None) -> list:
+    """
+    支持格式:
+    1. 视频文件目录
+    2. 单个视频文件
+    3. jsonl 文件: 每行 {"video", "question"[, "answer"]}
+    4. Video-R1 json 文件: [{"path", "problem", "options", "solution"}, ...]
+    """
+    path = Path(input_path)
+    items = []
+
+    if path.is_dir():
+        for f in sorted(path.rglob("*")):
+            if f.suffix.lower() in VIDEO_EXTENSIONS:
+                items.append({"video": str(f), "question": default_question})
+
+    elif path.is_file() and path.suffix.lower() in VIDEO_EXTENSIONS:
+        items.append({"video": str(path), "question": default_question})
+
+    elif path.suffix == ".jsonl":
+        with open(path, encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                item = json.loads(line)
+                item.setdefault("question", default_question)
+                items.append(item)
+
+    elif path.suffix == ".json":
+        with open(path, encoding="utf-8") as f:
+            data = json.load(f)
+        if not isinstance(data, list):
+            raise ValueError(f"JSON file should contain a list, got {type(data)}")
+        # 自动检测是否为 Video-R1 格式 (含 "path" 和 "problem" 字段)
+        if data and "path" in data[0] and "problem" in data[0]:
+            items = load_video_r1(input_path, video_root or ".", default_question)
+        else:
+            for item in data:
+                item.setdefault("question", default_question)
+                items.append(item)
+    else:
+        raise ValueError(f"Unsupported input format: {path.suffix}")
+
+    return items
+
+
+def load_processed(output_path: str) -> set:
+    """加载已处理的视频路径 (用于断点续跑)"""
+    processed = set()
+    if Path(output_path).exists():
+        with open(output_path, encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                entry = json.loads(line)
+                processed.add(entry.get("video", ""))
+    return processed
+
+
+# ============================================================
+# 处理流程
+# ============================================================
+def _make_result(item: dict) -> dict:
+    result = {"video": item["video"], "question": item["question"]}
+    if "answer" in item:
+        result["answer"] = item["answer"]
+    if "problem_id" in item:
+        result["problem_id"] = item["problem_id"]
+    return result
+
+
+def process_sequential(client, model, items, output_path, max_tokens, temperature, mode, max_frames, hint):
+    with open(output_path, mode, encoding="utf-8") as fout:
+        for item in tqdm(items, desc="Generating CoT"):
+            result = _make_result(item)
+            try:
+                result["response"] = generate_single(
+                    client, model,
+                    item["video"], item["question"],
+                    max_tokens, temperature,
+                    answer=item.get("answer"),
+                    max_frames=max_frames,
+                    hint=hint,
+                )
+            except Exception as e:
+                print(f"\n[ERROR] {item['video']}: {e}")
+                result["error"] = str(e)
+            fout.write(json.dumps(result, ensure_ascii=False) + "\n")
+            fout.flush()
+
+
+def process_parallel(client, model, items, output_path, max_tokens, temperature, mode, workers, max_frames, hint):
+    with open(output_path, mode, encoding="utf-8") as fout:
+        with ThreadPoolExecutor(max_workers=workers) as executor:
+            futures = {
+                executor.submit(
+                    generate_single,
+                    client, model,
+                    item["video"], item["question"],
+                    max_tokens, temperature,
+                    item.get("answer"),
+                    max_frames,
+                    hint,
+                ): item
+                for item in items
+            }
+            for future in tqdm(as_completed(futures), total=len(futures), desc="Generating CoT"):
+                item = futures[future]
+                result = _make_result(item)
+                try:
+                    result["response"] = future.result()
+                except Exception as e:
+                    print(f"\n[ERROR] {item['video']}: {e}")
+                    result["error"] = str(e)
+                fout.write(json.dumps(result, ensure_ascii=False) + "\n")
+                fout.flush()
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="批量视频 CoT 数据生成 (Qwen3.5 + ms-swift)",
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+    )
+    parser.add_argument("--input", required=True,
+                        help="输入: 视频目录 / jsonl 文件 / json 文件 (含 Video-R1 格式)")
+    parser.add_argument("--output", required=True,
+                        help="输出 jsonl 文件路径")
+    parser.add_argument("--video_root", default=None,
+                        help="Video-R1 json 格式时, 视频文件的根目录")
+    parser.add_argument("--api_base", default="http://localhost:8000/v1",
+                        help="ms-swift deploy 的 API 地址 (default: http://localhost:8000/v1)")
+    parser.add_argument("--model", default="Qwen3.5-122B-A10B",
+                        help="模型名称, 需与 swift deploy 一致")
+    parser.add_argument("--question", default=DEFAULT_QUESTION,
+                        help="默认问题 (当输入数据无 question 字段时使用)")
+    parser.add_argument("--max_tokens", type=int, default=4096,
+                        help="最大生成 token 数 (default: 4096)")
+    parser.add_argument("--temperature", type=float, default=0.7,
+                        help="生成温度 (default: 0.7)")
+    parser.add_argument("--workers", type=int, default=1,
+                        help="并行请求数 (default: 1, vLLM 会自动做 batch)")
+    parser.add_argument("--max_frames", type=int, default=None,
+                        help="最大采样帧数, 需与 FPS_MAX_FRAMES 环境变量一致 (默认读取环境变量, 否则 32)")
+    parser.add_argument("--hint", default="",
+                        help="维度专项提示，插入到 Question: 前，可直接传字符串或使用预设值: "
+                             "speed / direction / temporal")
+    parser.add_argument("--resume", action="store_true",
+                        help="断点续跑: 跳过已处理的视频")
+    args = parser.parse_args()
+
+    # hint 支持预设别名
+    _hint_map = {
+        "speed":     HINT_SPEED,
+        "direction": HINT_DIRECTION,
+        "temporal":  HINT_TEMPORAL,
+    }
+    hint = _hint_map.get(args.hint, args.hint)
+
+    # max_frames: 优先用 CLI 参数, 其次读环境变量 FPS_MAX_FRAMES, 最后默认 32
+    max_frames = args.max_frames or int(os.environ.get("FPS_MAX_FRAMES", 32))
+
+    items = load_inputs(args.input, args.question, args.video_root)
+    print(f"[INFO] Loaded {len(items)} items")
+    has_answer = sum(1 for i in items if "answer" in i)
+    print(f"[INFO] Items with ground-truth answer: {has_answer}/{len(items)}")
+    print(f"[INFO] Max sampled frames per video: {max_frames} (actual FPS varies by video duration)")
+
+    mode = "w"
+    if args.resume:
+        processed = load_processed(args.output)
+        before = len(items)
+        items = [item for item in items if item["video"] not in processed]
+        print(f"[INFO] Resume mode: skipped {before - len(items)}, remaining {len(items)}")
+        mode = "a"
+
+    if not items:
+        print("[INFO] Nothing to process. Done!")
+        return
+
+    Path(args.output).parent.mkdir(parents=True, exist_ok=True)
+
+    client = OpenAI(api_key="EMPTY", base_url=args.api_base)
+    print(f"[INFO] Connecting to {args.api_base} ...")
+    try:
+        models = client.models.list()
+        print(f"[INFO] Server ready. Available models: {[m.id for m in models.data]}")
+    except Exception as e:
+        print(f"[ERROR] Cannot connect to server: {e}")
+        print("[HINT] Please run 'bash start_server.sh' first.")
+        return
+
+    t0 = time.time()
+    if args.workers <= 1:
+        process_sequential(client, args.model, items, args.output,
+                           args.max_tokens, args.temperature, mode, max_frames, hint)
+    else:
+        process_parallel(client, args.model, items, args.output,
+                         args.max_tokens, args.temperature, mode, args.workers, max_frames, hint)
+
+    elapsed = time.time() - t0
+    print(f"\n[DONE] Results saved to {args.output}")
+    print(f"[DONE] Processed {len(items)} videos in {elapsed:.1f}s "
+          f"({elapsed / len(items):.1f}s per video)")
+
+
+if __name__ == "__main__":
+    main()
+
+
+# tmux send-keys -t cot_gen C-c 2>/dev/null; sleep 1; tmux send-keys -t cot_gen "cd /root/zwj/CoT_generate && python generate_video_cot.py --input /root/zwj/CoT_generate/output/need_rerun_v2.jsonl --output /root/zwj/CoT_generate/output/need_rerun_v2_cot.jsonl --max_frames 768 --workers 4 --resume" Enter

cot_generate/rl_data/08_clean_rl_data.py

@@ -0,0 +1,163 @@
+"""
+数据清洗脚本：从 07_oracle_results_nothink + 05_delta_frame_mapping 生成 RL 训练数据
+
+过滤策略：
+1. 非多选题（answer 不符合 A/A.xxx 格式）→ 丢弃
+2. oracle_results 为空（推理失败）→ 丢弃
+3. 单 pattern 条目（1个delta）→ 直接保留
+4. 多 pattern 条目（2+个delta）→ 过滤掉含异常 oracle_answer 的 pattern
+   异常定义：oracle_answer 不是单个字母 A-E
+"""
+
+import json
+from pathlib import Path
+
+DATA_DIR = Path(__file__).resolve().parent
+
+ORACLE_FILE = DATA_DIR / "07_oracle_results_nothink.json"
+DELTA_FILE  = DATA_DIR / "05_delta_frame_mapping.json"
+OUTPUT_FILE = DATA_DIR / "08_rl_training_data.json"
+
+VALID_ANSWERS = set("ABCDE")
+
+
+def is_multiple_choice(answer: str) -> bool:
+    """判断是否为标准多选题格式：A / A. xxx / A xxx"""
+    ans = answer.strip()
+    return (len(ans) == 1 and ans in VALID_ANSWERS) or \
+           (len(ans) >= 2 and ans[0] in VALID_ANSWERS and ans[1] in ". ")
+
+
+def is_valid_oracle_answer(answer: str) -> bool:
+    """oracle_answer 是否为合法的单字母选项"""
+    return answer.strip() in VALID_ANSWERS
+
+
+def main():
+    # 加载 delta frame mapping
+    with open(DELTA_FILE) as f:
+        delta_data = json.load(f)
+    delta_map = {m["problem_id"]: m for m in delta_data["mappings"]}
+    print(f"[Delta] 加载 {len(delta_map)} 条")
+
+    # 加载 oracle results，同时构建 oracle 查找表
+    with open(ORACLE_FILE) as f:
+        oracle_data = json.load(f)
+    oracle_results = oracle_data["results"]
+    print(f"[Oracle] 加载 {len(oracle_results)} 条")
+
+    # 构建查找表：(problem_id, mask_delta_id) -> oracle entry
+    oracle_map = {}
+    for r in oracle_results:
+        pid = r["problem_id"]
+        for o in r.get("oracle_results", []):
+            oracle_map[(pid, o["mask_delta_id"])] = o
+
+    stats = {
+        "total_problems": len(oracle_results),
+        "skipped_not_mc": 0,
+        "skipped_empty_oracle": 0,
+        "skipped_no_delta": 0,
+        "output_problems": 0,
+        "single_pattern_kept": 0,
+        "multi_pattern_kept": 0,
+        "multi_pattern_abnormal_removed": 0,
+        "total_mask_patterns": 0,
+        "oracle_correct_count": 0,
+    }
+
+    training_data = []
+
+    for problem in oracle_results:
+        pid = problem["problem_id"]
+
+        # 过滤1：非多选题
+        if not is_multiple_choice(problem["gt_answer"]):
+            stats["skipped_not_mc"] += 1
+            continue
+
+        # 过滤2：oracle_results 为空
+        if not problem["oracle_results"]:
+            stats["skipped_empty_oracle"] += 1
+            continue
+
+        # 过滤3：没有 delta mapping
+        delta_item = delta_map.get(pid)
+        if delta_item is None:
+            stats["skipped_no_delta"] += 1
+            continue
+
+        delta_mask_patterns = delta_item.get("mask_patterns", [])
+        n_patterns = len(delta_mask_patterns)
+
+        # 构建 mask_patterns
+        mask_patterns = []
+        for p in delta_mask_patterns:
+            mid = p["mask_delta_id"]
+            oracle_entry = oracle_map.get((pid, mid))
+            oracle_correct = oracle_entry["oracle_correct"] if oracle_entry else None
+
+            # 多 pattern 时，过滤掉异常 oracle_answer 的 pattern
+            if n_patterns > 1 and oracle_entry is not None:
+                if not is_valid_oracle_answer(oracle_entry["oracle_answer"]):
+                    stats["multi_pattern_abnormal_removed"] += 1
+                    continue
+
+            pattern = {
+                "mask_delta_id": mid,
+                "zero_frame_indices": p["zero_frame_indices"],
+                "remaining_frames": p["remaining_frames"],
+                "mask_ratio": p["mask_ratio"],
+                "oracle_correct": oracle_correct,
+            }
+            mask_patterns.append(pattern)
+            stats["total_mask_patterns"] += 1
+            if oracle_correct:
+                stats["oracle_correct_count"] += 1
+
+        # 过滤掉清洗后 mask_patterns 为空的条目
+        if not mask_patterns:
+            continue
+
+        entry = {
+            "problem_id": pid,
+            "video": problem["video_path"],
+            "question": problem["question"],
+            "answer": problem["gt_answer"],
+            "n_sampled_frames": problem["n_sampled_frames"],
+            "mask_patterns": mask_patterns,
+        }
+        training_data.append(entry)
+        stats["output_problems"] += 1
+
+        if n_patterns == 1:
+            stats["single_pattern_kept"] += 1
+        else:
+            stats["multi_pattern_kept"] += 1
+
+    # 写出
+    n_total_p = stats["total_mask_patterns"]
+    output = {
+        "metadata": {
+            "total_entries": len(training_data),
+            "total_mask_patterns": n_total_p,
+            "oracle_correct_count": stats["oracle_correct_count"],
+            "oracle_correct_rate": round(stats["oracle_correct_count"] / n_total_p, 3) if n_total_p else 0,
+            "oracle_source": ORACLE_FILE.name,
+            "delta_source": DELTA_FILE.name,
+        },
+        "data": training_data,
+    }
+
+    with open(OUTPUT_FILE, "w", encoding="utf-8") as f:
+        json.dump(output, f, ensure_ascii=False, indent=2)
+
+    print(f"\n{'='*50}")
+    print(f"清洗完成，输出到 {OUTPUT_FILE.name}")
+    print(f"{'='*50}")
+    for k, v in stats.items():
+        print(f"  {k}: {v}")
+
+
+if __name__ == "__main__":
+    main()

cot_generate/rl_data/09_convert_to_training_format.py

@@ -0,0 +1,184 @@
+"""
+格式转换：将 08_rl_training_data.json 转换为 DM-GRPO 训练所需格式
+
+目标格式（对齐 rl_delta_mask_training_data_clean.json）：
+{
+    "problem_id": 10,
+    "problem": "What are the first two ingredients shown in the video?",
+    "data_type": "video",
+    "problem_type": "multiple choice",
+    "options": ["A. Potato and garlic", "B. Potato and onion", ...],
+    "solution": "<answer>B</answer>",
+    "path": "./LLaVA-Video-178K/.../video.mp4",   # 相对路径
+    "data_source": "",
+    "video_duration": 32.7,
+    "n_sampled_frames": 16,
+    "mask_patterns": [...]
+}
+"""
+
+import json
+import re
+from pathlib import Path
+
+DATA_DIR = Path(__file__).resolve().parent
+INPUT_FILE  = DATA_DIR / "08_rl_training_data.json"
+DELTA_FILE  = DATA_DIR / "05_delta_frame_mapping.json"
+RAW_FILE    = DATA_DIR / "01_raw_input.jsonl"   # 完整 question 来源
+OUTPUT_FILE = DATA_DIR / "09_rl_training_data_formatted.json"
+
+# video 绝对路径的根前缀，转成相对路径时去掉
+VIDEO_ROOT = "/root/zwj/final_dataset/videos_timestamped"
+
+
+def parse_question_and_options(question_text: str):
+    """
+    把含选项的 question 拆分成 problem（纯题干）和 options 列表
+
+    输入示例：
+      "What are the first two ingredients?\nA. Potato and garlic\nB. Potato and onion\nC. ..."
+    输出：
+      problem = "What are the first two ingredients?"
+      options = ["A. Potato and garlic", "B. Potato and onion", ...]
+    """
+    lines = question_text.strip().split("\n")
+    option_pattern = re.compile(r"^[A-E][.、\s]")
+
+    problem_lines = []
+    options = []
+    in_options = False
+
+    for line in lines:
+        line = line.strip()
+        if not line:
+            continue
+        if option_pattern.match(line):
+            in_options = True
+            options.append(line)
+        elif not in_options:
+            problem_lines.append(line)
+
+    problem = " ".join(problem_lines).strip()
+    return problem, options
+
+
+def extract_answer_letter(answer: str) -> str:
+    """
+    从 gt_answer 中提取单字母选项
+
+    输入：'B. Potato and onion' 或 'B' 或 'B) xxx'
+    输出：'B'
+    """
+    ans = answer.strip()
+    if len(ans) == 1 and ans in "ABCDE":
+        return ans
+    if len(ans) >= 1 and ans[0] in "ABCDE":
+        return ans[0]
+    return ans
+
+
+def make_path_relative(video_abs_path: str) -> str:
+    """绝对路径转相对路径，去掉 VIDEO_ROOT 前缀"""
+    if video_abs_path.startswith(VIDEO_ROOT):
+        rel = video_abs_path[len(VIDEO_ROOT):]
+        # 保留 ./ 开头格式
+        return "." + rel
+    return video_abs_path
+
+
+def main():
+    # 加载清洗后数据
+    with open(INPUT_FILE) as f:
+        d = json.load(f)
+    input_data = d["data"]
+    print(f"[Input] 加载 {len(input_data)} 条")
+
+    # 加载原始数据，获取完整 question 和 data_source
+    raw_map = {}
+    with open(RAW_FILE) as f:
+        for line in f:
+            item = json.loads(line)
+            raw_map[item["problem_id"]] = item
+    print(f"[Raw] 加载 {len(raw_map)} 条原始数据")
+
+    # 加载 delta mapping 获取 video_duration
+    with open(DELTA_FILE) as f:
+        delta_raw = json.load(f)
+    duration_map = {
+        m["problem_id"]: m["video_info"]["duration"]
+        for m in delta_raw["mappings"]
+    }
+    print(f"[Delta] 加载 {len(duration_map)} 条 duration 信息")
+
+    stats = {
+        "total": len(input_data),
+        "missing_duration": 0,
+        "question_from_raw": 0,   # 用原始数据补全 question 的条数
+        "options_empty": 0,
+        "output": 0,
+    }
+
+    output_data = []
+
+    for item in input_data:
+        pid = item["problem_id"]
+
+        # 优先用原始数据的完整 question（防止截断）
+        raw_item = raw_map.get(pid)
+        question_text = raw_item["question"] if raw_item else item["question"]
+        if raw_item and question_text != item["question"]:
+            stats["question_from_raw"] += 1
+
+        # 拆分 problem 和 options
+        problem, options = parse_question_and_options(question_text)
+        if not options:
+            stats["options_empty"] += 1
+
+        # 提取答案字母
+        answer_letter = extract_answer_letter(item["answer"])
+        solution = f"<answer>{answer_letter}</answer>"
+
+        # 转换 video 路径
+        path = make_path_relative(item["video"])
+
+        # 获取 video_duration
+        duration = duration_map.get(pid)
+        if duration is None:
+            stats["missing_duration"] += 1
+            duration = 0.0
+
+        # data_source
+        data_source = raw_item["data_source"] if raw_item else ""
+
+        entry = {
+            "problem_id": pid,
+            "problem": problem,
+            "data_type": "video",
+            "problem_type": "multiple choice",
+            "options": options,
+            "solution": solution,
+            "path": path,
+            "data_source": data_source,
+            "video_duration": duration,
+            "n_sampled_frames": item["n_sampled_frames"],
+            "mask_patterns": item["mask_patterns"],
+        }
+        output_data.append(entry)
+        stats["output"] += 1
+
+    with open(OUTPUT_FILE, "w", encoding="utf-8") as f:
+        json.dump(output_data, f, ensure_ascii=False, indent=2)
+
+    print(f"\n{'='*50}")
+    print(f"转换完成，输出到 {OUTPUT_FILE.name}")
+    print(f"{'='*50}")
+    for k, v in stats.items():
+        print(f"  {k}: {v}")
+
+    # 打印一条示例对比
+    print("\n=== 输出示例 ===")
+    print(json.dumps(output_data[0], ensure_ascii=False, indent=2))
+
+
+if __name__ == "__main__":
+    main()

cot_generate/rl_data/build_delta_frame_mapping.py

@@ -0,0 +1,294 @@
+"""
+Build delta-to-frame-index mapping for Adaptive Delta Mask.
+
+For each sample:
+1. Parse CoT response to extract delta time ranges
+2. Load video metadata (total_frames, fps) via OpenCV
+3. Replicate Qwen2VL frame sampling logic (smart_nframes + linspace)
+4. Map each delta's time_range to sampled frame indices
+5. Generate mask patterns (which frames to zero out per delta)
+
+Usage:
+    python build_delta_frame_mapping.py \
+        --input Example/sampled_50.jsonl \
+        --output Example/delta_frame_mapping.json
+"""
+
+import argparse
+import json
+import math
+import os
+import re
+import sys
+import random
+
+import cv2
+
+# ---------------------------------------------------------------------------
+# Replicate Qwen2VL frame sampling logic from Video-R1
+# ---------------------------------------------------------------------------
+FRAME_FACTOR = 2
+DEFAULT_FPS = 2.0
+FPS_MIN_FRAMES = 4
+FPS_MAX_FRAMES_DEFAULT = 16
+
+
+def smart_nframes(total_frames: int, video_fps: float,
+                  sample_fps: float = DEFAULT_FPS,
+                  max_frames: int = FPS_MAX_FRAMES_DEFAULT) -> int:
+    """Replicate qwen_vl_utils.vision_process.smart_nframes."""
+    min_frames = math.ceil(FPS_MIN_FRAMES / FRAME_FACTOR) * FRAME_FACTOR
+    max_f = math.floor(min(max_frames, total_frames) / FRAME_FACTOR) * FRAME_FACTOR
+    nframes = total_frames / video_fps * sample_fps
+    nframes = min(min(max(nframes, min_frames), max_f), total_frames)
+    nframes = math.floor(nframes / FRAME_FACTOR) * FRAME_FACTOR
+    nframes = max(nframes, FRAME_FACTOR)
+    nframes = min(nframes, total_frames)
+    return int(nframes)
+
+
+def get_sampled_frame_info(video_path: str, max_frames: int = FPS_MAX_FRAMES_DEFAULT):
+    """Get video metadata and compute sampled frame indices + timestamps."""
+    cap = cv2.VideoCapture(video_path)
+    if not cap.isOpened():
+        return None
+
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    video_fps = cap.get(cv2.CAP_PROP_FPS)
+    cap.release()
+
+    if total_frames <= 0 or video_fps <= 0:
+        return None
+
+    duration = total_frames / video_fps
+    nframes = smart_nframes(total_frames, video_fps, max_frames=max_frames)
+
+    # Replicate torch.linspace(0, total_frames-1, nframes)
+    if nframes == 1:
+        indices = [0]
+    else:
+        step = (total_frames - 1) / (nframes - 1)
+        indices = [round(i * step) for i in range(nframes)]
+
+    # Each sampled frame's timestamp in seconds
+    timestamps = [idx / video_fps for idx in indices]
+
+    return {
+        "total_frames": total_frames,
+        "video_fps": video_fps,
+        "duration": duration,
+        "n_sampled": nframes,
+        "sampled_indices": indices,
+        "sampled_timestamps": timestamps,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Parse CoT response to extract deltas
+# ---------------------------------------------------------------------------
+def parse_timestamp(ts_str: str) -> float:
+    """Parse HH:MM:SS or HH:MM:SS.s to seconds."""
+    # Strip nested tags like <time>...</time>
+    ts_str = re.sub(r'<[^>]+>', '', ts_str).strip()
+    parts = ts_str.split(":")
+    if len(parts) == 3:
+        h, m, s = parts
+        return int(h) * 3600 + int(m) * 60 + float(s)
+    elif len(parts) == 2:
+        m, s = parts
+        return int(m) * 60 + float(s)
+    else:
+        return float(ts_str)
+
+
+def parse_deltas_from_cot(response: str):
+    """Extract delta time ranges from CoT response.
+
+    Looks for patterns like:
+        From <time_range>00:00:00 - 00:00:07</time_range>, <delta>...</delta>
+    """
+    time_ranges = re.findall(r'<time_range>(.*?)</time_range>', response)
+    deltas_text = re.findall(r'<delta>(.*?)</delta>', response, re.DOTALL)
+
+    deltas = []
+    for i, tr in enumerate(time_ranges):
+        try:
+            # Try regex first: match two timestamp-like groups separated by ' - '
+            m = re.match(r'\s*([\d:\.]+)\s*-\s*([\d:\.]+)\s*', re.sub(r'<[^>]+>', '', tr))
+            if not m:
+                continue
+            start = parse_timestamp(m.group(1))
+            end = parse_timestamp(m.group(2))
+            delta_text = deltas_text[i] if i < len(deltas_text) else ""
+            deltas.append({
+                "delta_id": i,
+                "time_range_str": tr.strip(),
+                "start_sec": start,
+                "end_sec": end,
+                "description": delta_text.strip(),
+            })
+        except (ValueError, AttributeError):
+            continue
+
+    return deltas
+
+
+# ---------------------------------------------------------------------------
+# Map deltas to sampled frame indices
+# ---------------------------------------------------------------------------
+def map_delta_to_frames(delta, sampled_timestamps):
+    """Find which sampled frame indices fall within the delta's time range.
+
+    Since CoT timestamps are integer seconds, we use floor-based matching:
+    a frame at timestamp t belongs to delta [start, end] if
+    int(t) is within [start_sec, end_sec].
+    """
+    start = int(delta["start_sec"])
+    end = int(delta["end_sec"])
+    frame_indices = []
+    for i, ts in enumerate(sampled_timestamps):
+        if start <= int(ts) <= end:
+            frame_indices.append(i)
+    return frame_indices
+
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+def process_entry(entry, video_root=None, max_frames=FPS_MAX_FRAMES_DEFAULT):
+    """Process a single data entry."""
+    video_path = entry["video"]
+    if video_root and not os.path.isabs(video_path):
+        video_path = os.path.join(video_root, video_path)
+
+    if not os.path.exists(video_path):
+        return {"problem_id": entry["problem_id"], "error": f"Video not found: {video_path}"}
+
+    # Get frame sampling info
+    frame_info = get_sampled_frame_info(video_path, max_frames=max_frames)
+    if frame_info is None:
+        return {"problem_id": entry["problem_id"], "error": "Failed to read video"}
+
+    # Parse deltas from CoT
+    deltas = parse_deltas_from_cot(entry["response"])
+    if not deltas:
+        return {"problem_id": entry["problem_id"], "error": "No deltas found in CoT"}
+
+    # Map each delta to frame indices
+    for d in deltas:
+        d["frame_indices"] = map_delta_to_frames(d, frame_info["sampled_timestamps"])
+
+    # Generate mask patterns: one pattern per delta (zero out that delta's frames)
+    mask_patterns = []
+    all_candidates = []  # 兜底用：保存所有候选 pattern
+    min_remaining = 4  # 至少保留 4 帧有效内容，避免信息过少产生噪声
+    for d in deltas:
+        if d["frame_indices"]:  # only if the delta maps to at least one frame
+            remaining = frame_info["n_sampled"] - len(d["frame_indices"])
+            pattern = {
+                "mask_delta_id": d["delta_id"],
+                "mask_delta_time": d["time_range_str"],
+                "zero_frame_indices": d["frame_indices"],
+                "remaining_frames": remaining,
+                "mask_ratio": round(len(d["frame_indices"]) / frame_info["n_sampled"], 3),
+            }
+            all_candidates.append(pattern)
+            if remaining >= min_remaining:
+                mask_patterns.append(pattern)
+
+    # 兜底：如果过滤后没有 pattern，保留剩余帧数最多的那个
+    if not mask_patterns and all_candidates:
+        best = max(all_candidates, key=lambda p: p["remaining_frames"])
+        mask_patterns.append(best)
+
+    result = {
+        "problem_id": entry["problem_id"],
+        "video_path": video_path,
+        "question": entry["question"],
+        "answer": entry.get("answer", ""),
+        "video_info": {
+            "total_frames": frame_info["total_frames"],
+            "video_fps": round(frame_info["video_fps"], 2),
+            "duration": round(frame_info["duration"], 2),
+            "n_sampled_frames": frame_info["n_sampled"],
+        },
+        "sampled_timestamps": [round(t, 2) for t in frame_info["sampled_timestamps"]],
+        "deltas": deltas,
+        "mask_patterns": mask_patterns,
+    }
+
+    return result
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Build delta-to-frame mapping")
+    parser.add_argument("--input", required=True, help="Input JSONL file with CoT data")
+    parser.add_argument("--output", required=True, help="Output JSON file")
+    parser.add_argument("--video-root", default=None, help="Root dir for relative video paths")
+    parser.add_argument("--max-frames", type=int, default=FPS_MAX_FRAMES_DEFAULT,
+                        help=f"Max sampled frames, must match RL training setting (default: {FPS_MAX_FRAMES_DEFAULT})")
+    args = parser.parse_args()
+
+    entries = []
+    with open(args.input) as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                entries.append(json.loads(line))
+
+    print(f"Processing {len(entries)} entries (max_frames={args.max_frames})...")
+
+    results = []
+    errors = []
+    for i, entry in enumerate(entries):
+        result = process_entry(entry, args.video_root, max_frames=args.max_frames)
+        if "error" in result:
+            errors.append(result)
+            print(f"  [{i+1}/{len(entries)}] ERROR {result['problem_id']}: {result['error']}")
+        else:
+            results.append(result)
+            n_deltas = len(result["deltas"])
+            n_patterns = len(result["mask_patterns"])
+            print(f"  [{i+1}/{len(entries)}] OK {result['problem_id']}: "
+                  f"{result['video_info']['n_sampled_frames']} frames, "
+                  f"{n_deltas} deltas, {n_patterns} mask patterns")
+
+    # Summary stats
+    print(f"\n{'='*60}")
+    print(f"Total: {len(entries)} | Success: {len(results)} | Errors: {len(errors)}")
+    if results:
+        avg_deltas = sum(len(r["deltas"]) for r in results) / len(results)
+        avg_patterns = sum(len(r["mask_patterns"]) for r in results) / len(results)
+        avg_frames = sum(r["video_info"]["n_sampled_frames"] for r in results) / len(results)
+        mask_ratios = [p["mask_ratio"] for r in results for p in r["mask_patterns"]]
+        avg_mask_ratio = sum(mask_ratios) / len(mask_ratios) if mask_ratios else 0
+        print(f"Avg sampled frames: {avg_frames:.1f}")
+        print(f"Avg deltas per video: {avg_deltas:.1f}")
+        print(f"Avg mask patterns: {avg_patterns:.1f}")
+        print(f"Avg mask ratio: {avg_mask_ratio:.3f} ({avg_mask_ratio*100:.1f}%)")
+
+    # Save
+    output_data = {
+        "metadata": {
+            "total_entries": len(entries),
+            "successful": len(results),
+            "errors": len(errors),
+            "frame_sampling": {
+                "method": "smart_nframes (Qwen2VL)",
+                "max_frames": args.max_frames,
+                "sample_fps": DEFAULT_FPS,
+            },
+            "mask_strategy": "delta-level zero-out",
+        },
+        "mappings": results,
+        "errors": errors,
+    }
+
+    os.makedirs(os.path.dirname(args.output) or ".", exist_ok=True)
+    with open(args.output, "w") as f:
+        json.dump(output_data, f, indent=2, ensure_ascii=False)
+    print(f"\nSaved to {args.output}")
+
+
+if __name__ == "__main__":
+    main()

cot_generate/rl_data/run_oracle_inference.py

@@ -0,0 +1,379 @@
+"""
+Oracle 推理脚本 — 对 masked 视频逐帧发送图片给强模型，获取 oracle 答案。
+
+核心逻辑:
+  1. 读取 delta_frame_mapping.json（含每条数据的所有 mask pattern）
+  2. 对每个 mask pattern:
+     - 用跟 trainer 完全一致的 smart_nframes 抽帧
+     - 将 delta 对应帧置零（黑帧）
+     - 逐帧编码为 base64 JPEG 图片
+     - 发给强模型 API（多图输入），获取回答
+     - 提取答案并与 GT 比对，记录 oracle_correct
+  3. 输出带 oracle 结果的 JSON
+
+使用方式:
+  1. 先启动 Qwen3.5 推理服务: bash start_server.sh
+  2. 运行:
+     python run_oracle_inference.py \
+       --mapping Example/delta_frame_mapping.json \
+       --output Example/oracle_results.json \
+       --max-frames 16
+
+依赖: cv2, numpy, openai, torch
+"""
+
+import argparse
+import base64
+import json
+import math
+import os
+import re
+import time
+import threading
+from concurrent.futures import ThreadPoolExecutor, as_completed
+from io import BytesIO
+
+import cv2
+import numpy as np
+from openai import OpenAI
+from tqdm import tqdm
+
+
+# ---------------------------------------------------------------------------
+# Frame sampling — 与 trainer (Qwen2VL) 完全一致
+# ---------------------------------------------------------------------------
+FRAME_FACTOR = 2
+DEFAULT_FPS = 2.0
+FPS_MIN_FRAMES = 4
+FPS_MAX_FRAMES_DEFAULT = 16
+
+
+def smart_nframes(total_frames: int, video_fps: float,
+                  sample_fps: float = DEFAULT_FPS,
+                  max_frames: int = FPS_MAX_FRAMES_DEFAULT) -> int:
+    """Replicate qwen_vl_utils.vision_process.smart_nframes."""
+    min_f = math.ceil(FPS_MIN_FRAMES / FRAME_FACTOR) * FRAME_FACTOR
+    max_f = math.floor(min(max_frames, total_frames) / FRAME_FACTOR) * FRAME_FACTOR
+    n = total_frames / video_fps * sample_fps
+    n = min(min(max(n, min_f), max_f), total_frames)
+    n = math.floor(n / FRAME_FACTOR) * FRAME_FACTOR
+    n = max(n, FRAME_FACTOR)
+    n = min(n, total_frames)
+    return int(n)
+
+
+def sample_frames(video_path: str, max_frames: int = FPS_MAX_FRAMES_DEFAULT):
+    """
+    用跟 trainer 一致的逻辑采样帧，返回 (frames_bgr_list, sampled_timestamps).
+    frames_bgr_list: list of numpy arrays (H, W, 3) in BGR
+    sampled_timestamps: list of float (seconds)
+    """
+    cap = cv2.VideoCapture(video_path)
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    video_fps = cap.get(cv2.CAP_PROP_FPS)
+
+    if total_frames <= 0 or video_fps <= 0:
+        cap.release()
+        raise ValueError(f"Cannot read video: {video_path}")
+
+    nframes = smart_nframes(total_frames, video_fps, max_frames=max_frames)
+
+    # linspace 采样 — 跟 torch.linspace(0, total_frames-1, nframes).round().long() 一致
+    indices = np.linspace(0, total_frames - 1, nframes).round().astype(int).tolist()
+    timestamps = [idx / video_fps for idx in indices]
+
+    frames = []
+    for idx in indices:
+        cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
+        ret, frame = cap.read()
+        if ret:
+            frames.append(frame)
+        else:
+            # fallback: 黑帧
+            h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) or 480
+            w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH)) or 640
+            frames.append(np.zeros((h, w, 3), dtype=np.uint8))
+    cap.release()
+
+    return frames, timestamps
+
+
+def apply_mask(frames, zero_frame_indices):
+    """将指定索引的帧置零（黑帧），返回新的 frames 列表。"""
+    masked = []
+    for i, f in enumerate(frames):
+        if i in set(zero_frame_indices):
+            masked.append(np.zeros_like(f))
+        else:
+            masked.append(f.copy())
+    return masked
+
+
+def frames_to_base64_content(frames, quality=85):
+    """将帧列表编码为 OpenAI API 多图格式的 content list。"""
+    content = []
+    for frame in frames:
+        _, buf = cv2.imencode(".jpg", frame, [cv2.IMWRITE_JPEG_QUALITY, quality])
+        b64 = base64.b64encode(buf.tobytes()).decode("utf-8")
+        content.append({
+            "type": "image_url",
+            "image_url": {"url": f"data:image/jpeg;base64,{b64}"}
+        })
+    return content
+
+
+# ---------------------------------------------------------------------------
+# Oracle prompt — 简洁直接，只要求回答问题
+# ---------------------------------------------------------------------------
+ORACLE_SYSTEM_PROMPT = """\
+You are a video understanding expert. You will be given a sequence of video frames \
+(images in temporal order) and a question. Analyze the frames carefully and answer \
+the question. First think about the reasoning process in <think> </think> tags, \
+then provide your final answer in <answer> </answer> tags, i.e., \
+<think> reasoning process here </think><answer> answer here </answer>
+
+Here is your question:
+{question}"""
+
+
+def extract_answer(text: str) -> str:
+    """从模型输出中提取 <answer>...</answer> 内容。"""
+    m = re.search(r"<answer>(.*?)</answer>", text, re.DOTALL)
+    if m:
+        return m.group(1).strip()
+    # fallback: 取最后一行非空文本
+    lines = [l.strip() for l in text.strip().split("\n") if l.strip()]
+    return lines[-1] if lines else text.strip()
+
+
+def normalize_answer(ans: str) -> str:
+    """标准化答案用于比较 — 取首字母/首词。"""
+    ans = ans.strip().rstrip(".")
+    # 处理 "A. xxx" 或 "A" 格式
+    m = re.match(r"^([A-D])\b", ans)
+    if m:
+        return m.group(1).upper()
+    return ans.lower()
+
+
+def check_correct(oracle_answer: str, gt_answer: str) -> bool:
+    """比较 oracle 答案与 GT 答案是否一致。"""
+    o = normalize_answer(oracle_answer)
+    g = normalize_answer(gt_answer)
+    # 直接匹配
+    if o == g:
+        return True
+    # GT 可能是 "A" 而 oracle 是 "A. xxx"
+    if len(g) == 1 and g.upper() in "ABCD":
+        return o == g.upper()
+    # oracle 可能是 "A" 而 GT 是 "A. xxx"
+    if len(o) == 1 and o.upper() in "ABCD":
+        return o.upper() == normalize_answer(gt_answer)
+    # 包含关系
+    return g in o or o in g
+
+
+# ---------------------------------------------------------------------------
+# API 调用
+# ---------------------------------------------------------------------------
+def call_oracle(client, model, frames_content, question, max_tokens=1024, temperature=0.0):
+    """调用强模型对 masked 帧序列进行推理。"""
+    system = ORACLE_SYSTEM_PROMPT.format(question=question)
+    user_content = frames_content + [
+        {"type": "text", "text": "These are video frames in temporal order. Please analyze them and answer the question."}
+    ]
+    response = client.chat.completions.create(
+        model=model,
+        messages=[
+            {"role": "system", "content": system},
+            {"role": "user", "content": user_content},
+        ],
+        max_tokens=max_tokens,
+        temperature=temperature,
+    )
+    return response.choices[0].message.content
+
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+def process_mapping(mapping_entry, client, model, max_frames, max_tokens, temperature, retry=2):
+    """处理一条 mapping entry 的所有 mask patterns。"""
+    video_path = mapping_entry["video_path"]
+    question = mapping_entry["question"]
+    gt_answer = mapping_entry.get("answer", "")
+
+    # 采帧（只采一次，所有 mask pattern 复用）
+    try:
+        frames, timestamps = sample_frames(video_path, max_frames=max_frames)
+    except Exception as e:
+        return {"problem_id": mapping_entry["problem_id"], "error": f"sample_frames: {e}"}
+
+    results = []
+    for pattern in mapping_entry.get("mask_patterns", []):
+        mask_indices = pattern["zero_frame_indices"]
+
+        # 置零
+        masked_frames = apply_mask(frames, mask_indices)
+        frames_content = frames_to_base64_content(masked_frames)
+
+        # 调用 API（带重试）
+        oracle_raw = None
+        for attempt in range(retry + 1):
+            try:
+                oracle_raw = call_oracle(client, model, frames_content, question,
+                                         max_tokens=max_tokens, temperature=temperature)
+                break
+            except Exception as e:
+                if attempt < retry:
+                    time.sleep(2 ** attempt)
+                else:
+                    oracle_raw = f"[ERROR] {e}"
+
+        oracle_ans = extract_answer(oracle_raw) if not oracle_raw.startswith("[ERROR]") else ""
+        correct = check_correct(oracle_ans, gt_answer) if oracle_ans else False
+
+        results.append({
+            "mask_delta_id": pattern["mask_delta_id"],
+            "mask_delta_time": pattern["mask_delta_time"],
+            "zero_frame_indices": mask_indices,
+            "mask_ratio": pattern["mask_ratio"],
+            "oracle_raw_response": oracle_raw,
+            "oracle_answer": oracle_ans,
+            "gt_answer": gt_answer,
+            "oracle_correct": correct,
+        })
+
+    return {
+        "problem_id": mapping_entry["problem_id"],
+        "video_path": video_path,
+        "question": question[:100] + "..." if len(question) > 100 else question,
+        "gt_answer": gt_answer,
+        "n_sampled_frames": len(frames),
+        "oracle_results": results,
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Oracle 推理: 对 masked 视频逐帧发图片给强模型",
+    )
+    parser.add_argument("--mapping", required=True,
+                        help="delta_frame_mapping.json 文件路径")
+    parser.add_argument("--output", required=True,
+                        help="输出 JSON 文件路径")
+    parser.add_argument("--max-frames", type=int, default=FPS_MAX_FRAMES_DEFAULT,
+                        help=f"最大采样帧数，需与 RL 训练一致 (default: {FPS_MAX_FRAMES_DEFAULT})")
+    parser.add_argument("--api-base", default="http://localhost:8000/v1",
+                        help="强模型 API 地��� (default: http://localhost:8000/v1)")
+    parser.add_argument("--model", default="Qwen3.5-122B-A10B",
+                        help="强模型名称 (default: Qwen3.5-122B-A10B)")
+    parser.add_argument("--max-tokens", type=int, default=1024,
+                        help="最大生成 token 数 (default: 1024)")
+    parser.add_argument("--temperature", type=float, default=0.0,
+                        help="生成温度 (default: 0.0)")
+    parser.add_argument("--resume", action="store_true",
+                        help="断点续跑：跳过已有结果的 problem_id")
+    parser.add_argument("--workers", type=int, default=8,
+                        help="并发线程数 (default: 8)")
+    args = parser.parse_args()
+
+    # 加载 mapping
+    with open(args.mapping) as f:
+        data = json.load(f)
+    mappings = data["mappings"]
+    print(f"Loaded {len(mappings)} entries from {args.mapping}")
+    print(f"Settings: max_frames={args.max_frames}, model={args.model}, workers={args.workers}")
+
+    # 断点续跑：加载已完成的 problem_id
+    done_ids = set()
+    existing_results = []
+    if args.resume and os.path.exists(args.output):
+        with open(args.output) as f:
+            prev = json.load(f)
+        existing_results = prev.get("results", [])
+        done_ids = {r["problem_id"] for r in existing_results if "error" not in r}
+        print(f"Resume: skipping {len(done_ids)} already completed entries")
+
+    remaining = [m for m in mappings if m["problem_id"] not in done_ids]
+    print(f"To process: {len(remaining)} entries")
+
+    if not remaining:
+        print("Nothing to do.")
+        return
+
+    # 初始化 API client
+    client = OpenAI(base_url=args.api_base, api_key="not-needed")
+
+    # 并发处理
+    results = list(existing_results)
+    errors = []
+    lock = threading.Lock()
+    save_counter = [0]  # mutable counter for save interval
+
+    def _process_one(entry):
+        return process_mapping(entry, client, args.model,
+                               max_frames=args.max_frames,
+                               max_tokens=args.max_tokens,
+                               temperature=args.temperature)
+
+    pbar = tqdm(total=len(remaining), desc="Oracle inference")
+
+    with ThreadPoolExecutor(max_workers=args.workers) as executor:
+        futures = {executor.submit(_process_one, entry): entry for entry in remaining}
+
+        for future in as_completed(futures):
+            result = future.result()
+            with lock:
+                if "error" in result:
+                    errors.append(result)
+                    pbar.write(f"  ERROR {result['problem_id']}: {result['error']}")
+                else:
+                    n_correct = sum(1 for r in result["oracle_results"] if r["oracle_correct"])
+                    n_total = len(result["oracle_results"])
+                    pbar.write(f"  OK {result['problem_id']}: "
+                               f"{n_correct}/{n_total} oracle correct")
+                results.append(result)
+                save_counter[0] += 1
+                pbar.update(1)
+
+                # 每 10 条自动保存
+                if save_counter[0] % 10 == 0:
+                    _save(args.output, results, errors, args)
+
+    pbar.close()
+    _save(args.output, results, errors, args)
+
+    # 统计
+    all_oracle = []
+    for r in results:
+        if "oracle_results" in r:
+            all_oracle.extend(r["oracle_results"])
+
+    n_oracle_correct = sum(1 for o in all_oracle if o.get("oracle_correct"))
+    print(f"\n{'='*60}")
+    print(f"Total entries: {len(results)} | Errors: {len(errors)}")
+    print(f"Total mask patterns: {len(all_oracle)}")
+    print(f"Oracle correct: {n_oracle_correct}/{len(all_oracle)} "
+          f"({n_oracle_correct/len(all_oracle)*100:.1f}%)" if all_oracle else "")
+    print(f"Saved to {args.output}")
+
+
+def _save(output_path, results, errors, args):
+    """保存中间/最终结果。"""
+    output = {
+        "metadata": {
+            "total_entries": len(results),
+            "errors": len(errors),
+            "max_frames": args.max_frames,
+            "model": args.model,
+            "temperature": args.temperature,
+        },
+        "results": results,
+    }
+    with open(output_path, "w", encoding="utf-8") as f:
+        json.dump(output, f, ensure_ascii=False, indent=2)
+
+
+if __name__ == "__main__":
+    main()

cot_generate/rl_data/select_single_pattern.py

@@ -0,0 +1,73 @@
+"""
+从 full_delta_frame_mapping.json 中为每个视频随机选取一个 mask pattern，
+输出精简版 mapping，用于加速 oracle 推理。
+
+随机性保证:
+  - 使用 os.urandom 生成种子（系统级熵源）
+  - 每条 entry 独立随机选取
+  - 打印种子方便复现
+
+使用方式:
+  python select_single_pattern.py \
+    --input full_delta_frame_mapping.json \
+    --output single_delta_frame_mapping.json
+"""
+
+import argparse
+import json
+import os
+import random
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="为每个视频随机选取一个 mask pattern"
+    )
+    parser.add_argument("--input", required=True, help="完整 mapping JSON")
+    parser.add_argument("--output", required=True, help="输出精简 mapping JSON")
+    parser.add_argument("--seed", type=int, default=None,
+                        help="随机种子（不指定则用 os.urandom 自动生成）")
+    args = parser.parse_args()
+
+    # 随机种子：优先用户指定，否则用系统熵源
+    if args.seed is not None:
+        seed = args.seed
+    else:
+        seed = int.from_bytes(os.urandom(8), "big")
+    random.seed(seed)
+    print(f"Random seed: {seed}  (可用 --seed {seed} 复现)")
+
+    with open(args.input) as f:
+        data = json.load(f)
+
+    mappings = data["mappings"]
+    print(f"输入: {len(mappings)} entries")
+
+    total_before = 0
+    total_after = 0
+    no_pattern = 0
+
+    for entry in mappings:
+        patterns = entry.get("mask_patterns", [])
+        total_before += len(patterns)
+        if patterns:
+            selected = random.choice(patterns)
+            entry["mask_patterns"] = [selected]
+            total_after += 1
+        else:
+            no_pattern += 1
+
+    data["metadata"]["selection"] = "single_random_pattern"
+    data["metadata"]["random_seed"] = seed
+
+    with open(args.output, "w", encoding="utf-8") as f:
+        json.dump(data, f, ensure_ascii=False, indent=2)
+
+    print(f"输出: {len(mappings)} entries")
+    print(f"Patterns: {total_before} → {total_after} ({total_before - total_after} 减少)")
+    print(f"无 pattern 的 entry: {no_pattern}")
+    print(f"Saved to {args.output}")
+
+
+if __name__ == "__main__":
+    main()

cot_generate/run_generate.sh

@@ -0,0 +1,58 @@
+#!/bin/bash
+# ============================================================
+# 视频 CoT 数据生成 - 完整流程示例
+#
+# 流程:
+#   Step 1. 给视频加时间戳水印 (add_timestamp_to_video.py)
+#   Step 2. 启动 Qwen3.5 推理服务 (start_server.sh)
+#   Step 3. 批量生成 CoT 数据 (generate_video_cot.py)
+# ============================================================
+
+set -e
+cd "$(dirname "$0")"
+
+# ==============================
+# Step 1: 给视频加时间戳 (可选, 如果视频已有时间戳可跳过)
+# ==============================
+# python add_timestamp_to_video.py \
+#     --input_dir /path/to/raw_videos \
+#     --output_dir /path/to/timestamped_videos \
+#     --ext mp4
+
+# ==============================
+# Step 2: 启动推理服务 (在另一个终端运行, 等待服务就绪)
+# ==============================
+# bash start_server.sh
+# 等待看到类似 "Uvicorn running on http://0.0.0.0:8000" 的日志
+
+# ==============================
+# Step 3: 批量生成 CoT
+# ==============================
+
+# --- 示例 A: 输入为视频目录, 使用默认问题 ---
+# python generate_video_cot.py \
+#     --input /path/to/timestamped_videos \
+#     --output ./output/cot_results.jsonl
+
+# --- 示例 B: 输入为 jsonl, 每个视频有不同的问题 ---
+# python generate_video_cot.py \
+#     --input ./my_data.jsonl \
+#     --output ./output/cot_results.jsonl
+
+# --- 示例 C: 自定义问题 + 断点续跑 ---
+# python generate_video_cot.py \
+#     --input /path/to/timestamped_videos \
+#     --output ./output/cot_results.jsonl \
+#     --question "What is the velocity of the object in this video? Analyze its motion frame by frame." \
+#     --max_tokens 4096 \
+#     --temperature 0.7 \
+#     --resume
+
+# --- 示例 D: 实际运行示例 ---
+python generate_video_cot.py \
+    --input /path/to/your/videos \
+    --output ./output/cot_results.jsonl \
+    --question "Describe and analyze what is happening in this video step by step." \
+    --max_tokens 4096 \
+    --temperature 0.7 \
+    --resume

cot_generate/sft_data/filter_cot.py

@@ -0,0 +1,400 @@
+#!/usr/bin/env python3
+"""
+CoT Data Filtering Script
+=========================
+cot_all.jsonl -> SFT-ready JSON for sft_video_v5e.py
+
+Filter rules:
+  Structural:
+  1.  has 'error' key
+  2.  response empty or too short (< MIN_RESPONSE_LEN)
+  3.  <think> opened but never closed (truncated)
+  4.  <think> content too short (< MIN_THINK_LEN)
+  5.  no <time> / <time_range> tag
+  6.  no <reasoning_summary> tag
+  7.  </think> tail truncated or empty
+  8.  video file missing on disk
+  9.  no <answer> tag, or answer content empty
+  10. no <delta> tag
+  11. no "Reference frame at"
+  Quality:
+  12. meta language ("as an AI", "I cannot process", ...)
+  13. answer leakage ("answer key", "ground truth", ...)
+  14. MC: pred letter unparsable (e.g. placeholder "concise answer")
+  15. MC: pred letter != GT letter
+
+Problem type (dual condition based on GT answer):
+  question has option lines  AND  GT answer resolves to A-E letter -> multiple-choice
+  otherwise -> free-form (even if question has options but GT is free text)
+
+Output fields: problem_id, problem, data_type, problem_type, options,
+               process, solution, path, data_source
+"""
+
+import json
+import os
+import re
+import argparse
+
+MIN_THINK_LEN    = 200
+MIN_RESPONSE_LEN = 100
+
+# 原始无时间戳视频的文件名索引（文件名 → 路径列表）
+_VIDEO_INDEX_PATH = os.path.join(os.path.dirname(__file__), "orig_video_index.json")
+if os.path.exists(_VIDEO_INDEX_PATH):
+    with open(_VIDEO_INDEX_PATH) as _f:
+        _VIDEO_INDEX = json.load(_f)
+else:
+    _VIDEO_INDEX = {}
+
+META_KEYWORDS = [
+    "as an AI", "as a language model", "I apologize for",
+    "I'm unable to", "I cannot process", "I give up",
+    "for the sake of the task", "the prompt has a typo",
+]
+
+ANSWER_LEAK_PATTERNS = [
+    r'answer key says', r'answer key', r'the provided answer',
+    r'correct answer (?:is )?(?:given|provided|stated)',
+    r'answer is given as', r'ground truth', r'reference answer',
+    r'per the provided answer', r'as per the provided',
+    r'need to box [A-E].*?per the',
+]
+
+
+# ---------------------------------------------------------------------------
+# Helpers
+# ---------------------------------------------------------------------------
+
+def extract_think(response):
+    if "<think>" not in response or "</think>" not in response:
+        return None
+    return response.split("<think>", 1)[1].split("</think>", 1)[0]
+
+
+def extract_after_think(response):
+    if "</think>" in response:
+        return response.split("</think>", 1)[1].strip()
+    return response.strip()
+
+
+def extract_letter(text):
+    """Extract option letter A-E from answer text. Returns None on failure.
+    Supports: A / A. / A. text / (A) / (A) text
+    """
+    if not text:
+        return None
+    t = text.strip()
+    if re.match(r'^[A-E]$', t):
+        return t.upper()
+    m = re.match(r'^[(\u0028\uff08]?\s*([A-E])\s*[)\u0029\uff09.]', t)
+    if m:
+        return m.group(1).upper()
+    return None
+
+
+# 选项后面追加的指令模板，需要从 question 和 options 中剔除
+_TRAILING_INSTRUCTIONS = re.compile(
+    r'\n?(?:Please respond.*|Note:.*|Answer by.*|Provide your answer.*)$',
+    re.IGNORECASE | re.DOTALL
+)
+
+
+def strip_trailing_instructions(text):
+    """去掉 question 末尾的指令文字（如 'Please respond systematically...'）"""
+    return _TRAILING_INSTRUCTIONS.sub('', text).strip()
+
+
+def extract_options_from_question(question):
+    """Return [(letter, text), ...] for option lines in question.
+    Supports: A. text  A) text  (A) text
+    先去掉尾部指令再提取，避免最后一个选项文字混入指令。
+    """
+    q = strip_trailing_instructions(question)
+    pat = re.compile(
+        r'(?:^|\n)[(\u0028\uff08]?([A-E])[)\u0029\uff09.]\s*(.+?)(?=\n[(\u0028\uff08]?[A-E][)\u0029\uff09.]|\Z)',
+        re.DOTALL
+    )
+    return [(l, t.strip()) for l, t in pat.findall(q)]
+
+
+def classify_entry(question, gt_answer):
+    """Return (problem_type, options_list).
+    multiple choice only when BOTH: question has option lines AND GT has A-E letter.
+    Uses 'multiple choice' (space) to match sft_video_v5e.py TYPE_TEMPLATE keys.
+    """
+    opts = extract_options_from_question(question)
+    gt_letter = extract_letter(gt_answer)
+    if opts and gt_letter:
+        return "multiple choice", [f"{l}. {t.strip()}" for l, t in opts]
+    return "free-form", []
+
+
+def check_mc_answer(pred, gt_letter, question):
+    """Return (ok: bool, reason: str).
+    1. Try direct letter extraction.
+    2. Fallback: full option text match (model answered the text not the letter).
+    3. Still None -> unparsable.
+    """
+    pred_letter = extract_letter(pred)
+    if pred_letter is None:
+        opts = extract_options_from_question(question)
+        pred_clean = pred.lower().strip().rstrip(".")
+        for letter, opt_text in opts:
+            opt_clean = opt_text.strip().lower().rstrip(".")
+            if pred_clean == opt_clean or pred_clean in opt_clean or opt_clean in pred_clean:
+                pred_letter = letter.upper()
+                break
+    if pred_letter is None:
+        return False, "mc_unparsable"
+    if pred_letter != gt_letter:
+        return False, "mc_wrong_answer"
+    return True, "ok"
+
+
+def extract_structured_body(response):
+    """提取 </think> 之后、<answer> 之前的结构化内容
+    （Reference frame / delta / reasoning_summary 部分），
+    作为新的 <think> 块内容。
+    """
+    after = extract_after_think(response)
+    # 去掉末尾的 <answer>...</answer>
+    after = re.sub(r'<answer>.*?</answer>\s*$', '', after, flags=re.DOTALL).strip()
+    return after
+
+
+def build_output_entry(obj, video_data_root):
+    video_raw  = obj["video"]
+    answer     = obj.get("answer", "")
+    question   = obj.get("question", "")
+    response   = obj.get("response", "")
+    problem_id = obj.get("problem_id", "")
+
+    structured = extract_structured_body(response)
+    process  = f"<think>\n{structured}\n</think>"
+
+    # MC solution: 训练脚本期望只有字母（TYPE_TEMPLATE 里要求 "single option letter"）
+    # free-form solution: 保留完整 GT 文本
+    gt_letter = extract_letter(answer)
+    has_opts  = bool(extract_options_from_question(question))
+    if gt_letter and has_opts:
+        solution = f"<answer>{gt_letter}</answer>"
+    else:
+        solution = f"<answer>{answer.strip()}</answer>"
+
+    # path：优先使用原始无时间戳视频，回退顺序：
+    #   1. /final_dataset/<rel>（去掉 videos_timestamped 层）
+    #   2. /root/zwj/<rel>（同上，另一种前缀）
+    #   3. data_preparation/<subdir>/<fname>
+    #   4. compose_data/<subdir>/<fname>
+    #   5. 实在找不到则保留 timestamped 路径（如 *_reverse_*.mp4）
+    path = video_raw
+    ts_marker = "/final_dataset/videos_timestamped/"
+    if ts_marker in video_raw:
+        rel      = video_raw.split(ts_marker, 1)[1]          # e.g. dataset_First_Version/...
+        cand1    = os.path.join("/root/zwj/final_dataset", rel)
+        cand2    = os.path.join("/root/zwj", rel)
+        if os.path.exists(cand1):
+            path = cand1
+        elif os.path.exists(cand2):
+            path = cand2
+        else:
+            # 按文件名在索引中查找，优先级：data_preparation > compose_data > Video-R1
+            fname   = os.path.basename(video_raw)
+            matches = _VIDEO_INDEX.get(fname, [])
+            chosen  = None
+            for priority in ["data_preparation", "compose_data", "Video-R1"]:
+                for m in matches:
+                    if priority in m:
+                        chosen = m
+                        break
+                if chosen:
+                    break
+            if chosen:
+                path = chosen
+            # else: 保留原始 timestamped 路径
+
+    problem_type, options = classify_entry(question, answer)
+
+    # problem：先去掉尾部指令，再去掉选项块，只保留题目主体
+    q_clean = strip_trailing_instructions(question)
+    problem = re.sub(
+        r"\n?(?:Options?[:\s]*\n)?(?:[(\uff08]?[A-E][)\uff09.].+(?:\n|$))+",
+        "", q_clean, flags=re.IGNORECASE
+    ).strip()
+
+    return {
+        "problem_id":   problem_id,
+        "problem":      problem,
+        "data_type":    "video",
+        "problem_type": problem_type,
+        "options":      options,
+        "process":      process,
+        "solution":     solution,
+        "path":         path,
+        "data_source":  "cot_all_filtered",
+    }
+
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+
+def filter_and_convert(input_path, output_path, video_data_root):
+    reject_reasons = {
+        "error_key":                0,
+        "empty_response":           0,
+        "truncated_no_close_think": 0,
+        "short_think":              0,
+        "no_time_tag":              0,
+        "no_reasoning_summary":     0,
+        "truncated_after_think":    0,
+        "video_missing":            0,
+        "no_answer_tag":            0,
+        "no_delta":                 0,
+        "no_reference_frame":       0,
+        "meta_language":            0,
+        "answer_leak":              0,
+        "mc_unparsable":            0,
+        "mc_wrong_answer":          0,
+    }
+
+    total    = 0
+    passed   = []
+    rejected = []   # 保存被过滤的原始条目，附带过滤原因
+
+    with open(input_path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                obj = json.loads(line)
+            except json.JSONDecodeError:
+                continue
+
+            total += 1
+            resp     = obj.get("response", "")
+            question = obj.get("question", "")
+            gt       = obj.get("answer", "").strip()
+
+            # --- Structural filters ---
+
+            if "error" in obj:
+                reject_reasons["error_key"] += 1
+                rejected.append({**obj, "_reject_reason": "error_key"})
+                continue
+
+            if not resp or len(resp) < MIN_RESPONSE_LEN:
+                reject_reasons["empty_response"] += 1
+                rejected.append({**obj, "_reject_reason": "empty_response"})
+                continue
+
+            if "<think>" in resp and "</think>" not in resp:
+                reject_reasons["truncated_no_close_think"] += 1
+                rejected.append({**obj, "_reject_reason": "truncated_no_close_think"})
+                continue
+
+            think_body = extract_think(resp)
+            if think_body is None or len(think_body.strip()) < MIN_THINK_LEN:
+                reject_reasons["short_think"] += 1
+                rejected.append({**obj, "_reject_reason": "short_think"})
+                continue
+
+            if "<time>" not in resp and "<time_range>" not in resp:
+                reject_reasons["no_time_tag"] += 1
+                rejected.append({**obj, "_reject_reason": "no_time_tag"})
+                continue
+
+            if "<reasoning_summary>" not in resp:
+                reject_reasons["no_reasoning_summary"] += 1
+                rejected.append({**obj, "_reject_reason": "no_reasoning_summary"})
+                continue
+
+            after = extract_after_think(resp)
+            has_ans_tag = "<answer>" in after or "</answer>" in after
+            if len(after) < 10 or (after[-1] not in ".!?>" and not has_ans_tag):
+                reject_reasons["truncated_after_think"] += 1
+                rejected.append({**obj, "_reject_reason": "truncated_after_think"})
+                continue
+
+            video = obj.get("video", "")
+            if video and not os.path.exists(video):
+                reject_reasons["video_missing"] += 1
+                rejected.append({**obj, "_reject_reason": "video_missing"})
+                continue
+
+            ans_m = re.search(r'<answer>(.*?)</answer>', resp, re.DOTALL)
+            pred  = ans_m.group(1).strip() if ans_m else None
+            if not pred:
+                reject_reasons["no_answer_tag"] += 1
+                rejected.append({**obj, "_reject_reason": "no_answer_tag"})
+                continue
+
+            if "<delta>" not in resp:
+                reject_reasons["no_delta"] += 1
+                rejected.append({**obj, "_reject_reason": "no_delta"})
+                continue
+
+            if "Reference frame at" not in resp:
+                reject_reasons["no_reference_frame"] += 1
+                rejected.append({**obj, "_reject_reason": "no_reference_frame"})
+                continue
+
+            # --- Quality filters ---
+
+            lower_resp = resp.lower()
+            if any(kw.lower() in lower_resp for kw in META_KEYWORDS):
+                reject_reasons["meta_language"] += 1
+                rejected.append({**obj, "_reject_reason": "meta_language"})
+                continue
+
+            if any(re.search(p, resp, re.IGNORECASE) for p in ANSWER_LEAK_PATTERNS):
+                reject_reasons["answer_leak"] += 1
+                rejected.append({**obj, "_reject_reason": "answer_leak"})
+                continue
+
+            # MC answer correctness
+            gt_letter = extract_letter(gt)
+            has_opts  = bool(extract_options_from_question(question))
+            if gt_letter and has_opts:
+                ok, reason = check_mc_answer(pred, gt_letter, question)
+                if not ok:
+                    reject_reasons[reason] += 1
+                    rejected.append({**obj, "_reject_reason": reason})
+                    continue
+
+            passed.append(build_output_entry(obj, video_data_root))
+
+    with open(output_path, "w", encoding="utf-8") as f:
+        json.dump(passed, f, ensure_ascii=False, indent=2)
+
+    # 保存 rejected 条目（原始字段 + _reject_reason）
+    rejected_path = output_path.replace(".json", "_rejected.json")
+    with open(rejected_path, "w", encoding="utf-8") as f:
+        json.dump(rejected, f, ensure_ascii=False, indent=2)
+
+    rejected_cnt = sum(reject_reasons.values())
+    print("=" * 62)
+    print(f"  Input total  : {total:>7,}")
+    print(f"  Passed       : {len(passed):>7,}  ({100*len(passed)/total:.1f}%)")
+    print(f"  Rejected     : {rejected_cnt:>7,}  ({100*rejected_cnt/total:.1f}%)")
+    print("-" * 62)
+    for reason, cnt in reject_reasons.items():
+        if cnt > 0:
+            print(f"    {reason:<35}: {cnt:>5,}  ({100*cnt/total:.1f}%)")
+    print("=" * 62)
+    print(f"  Output:          {output_path}")
+    print(f"  Rejected saved:  {rejected_path}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--input",           default=None, help="Path to cot_all.jsonl")
+    parser.add_argument("--output",          default=None, help="Path to output filtered json")
+    parser.add_argument("--video_data_root", default=os.environ.get("VIDEO_DATA_ROOT", "/root/zwj"))
+    parser.add_argument("--min_think_len",   type=int, default=MIN_THINK_LEN)
+    args = parser.parse_args()
+
+    MIN_THINK_LEN = args.min_think_len
+    filter_and_convert(args.input, args.output, args.video_data_root)