Datasets:

ArXiv:
parap1uie-s commited on
Commit
c3dc892
·
1 Parent(s): 1e9e39f

fix video track docs and readme

Browse files
video_understanding/README.md CHANGED
@@ -20,7 +20,7 @@ Each video is a handheld recording of credit-review documents (2—4 per clip) a
20
  ```
21
  FCMBench-Video_v1.0_Videos/
22
  └── video/
23
- ├── Construction/ # CN original videos (15 persons × 9 = 135 clips)
24
  ├── Construction-en-US/ # EN original videos (30 persons × 9 = 265 clips)
25
  ├── VPI-cn/ # CN videos with visual prompt injection (135 clips)
26
  ├── VPI-cot-cn/ # CN VPI + Chain-of-Thought (135 clips)
@@ -34,7 +34,6 @@ FCMBench-Video_v1.0_Videos/
34
  |------|-------------|--------|----------|
35
  | `FCMBench-Video_v1.0_full.jsonl` | Full 11,322 samples (CN + EN merged) | ✗ | Public release; run inference |
36
  | `FCMBench-Video_v1.0_small.jsonl` | Stratified 10% sample (~1,135) | ✓ | Quick self-evaluation / sanity check |
37
- | `FCMBench-Video_v1.0_full-gt.jsonl` | Full 11,322 samples | ✓ | Internal reference (not distributed) |
38
 
39
  **Tasks** (7 types across perception & reasoning):
40
 
@@ -80,7 +79,7 @@ The unzipped tree will look like:
80
  ```
81
  FCMBench-Video_v1.0_Videos/
82
  └── video/
83
- ├── Construction/
84
  ├── Construction-en-US/
85
  ├── VPI-cn/
86
  ├── VPI-cot-cn/
@@ -89,7 +88,7 @@ FCMBench-Video_v1.0_Videos/
89
  ```
90
 
91
  The instruction JSONL files use `video_prefix` + `video_path` fields (e.g.
92
- `"video_prefix": "Construction", "video_path": "yangyimiao/yangyimiao_20s_1.mp4"`),
93
  so point `--video_root` to the `FCMBench-Video_v1.0_Videos/` directory.
94
 
95
  ### 2) Run inference
@@ -167,5 +166,4 @@ To have your model ranked on the **FCMBench-Video leaderboard**:
167
  - Inference framework (or API) and key settings (e.g., temperature, max tokens)
168
  - Any special post-processing applied (if applicable)
169
 
170
- After validation, we will compute the official metrics on the hidden ground-truth
171
- (`FCMBench-Video_v1.0_full-gt.jsonl`) and update the leaderboard.
 
20
  ```
21
  FCMBench-Video_v1.0_Videos/
22
  └── video/
23
+ ├── Construction-cn/ # CN original videos (15 persons × 9 = 135 clips)
24
  ├── Construction-en-US/ # EN original videos (30 persons × 9 = 265 clips)
25
  ├── VPI-cn/ # CN videos with visual prompt injection (135 clips)
26
  ├── VPI-cot-cn/ # CN VPI + Chain-of-Thought (135 clips)
 
34
  |------|-------------|--------|----------|
35
  | `FCMBench-Video_v1.0_full.jsonl` | Full 11,322 samples (CN + EN merged) | ✗ | Public release; run inference |
36
  | `FCMBench-Video_v1.0_small.jsonl` | Stratified 10% sample (~1,135) | ✓ | Quick self-evaluation / sanity check |
 
37
 
38
  **Tasks** (7 types across perception & reasoning):
39
 
 
79
  ```
80
  FCMBench-Video_v1.0_Videos/
81
  └── video/
82
+ ├── Construction-cn/
83
  ├── Construction-en-US/
84
  ├── VPI-cn/
85
  ├── VPI-cot-cn/
 
88
  ```
89
 
90
  The instruction JSONL files use `video_prefix` + `video_path` fields (e.g.
91
+ `"video_prefix": "Construction-cn", "video_path": "yangyimiao/yangyimiao_20s_1.mp4"`),
92
  so point `--video_root` to the `FCMBench-Video_v1.0_Videos/` directory.
93
 
94
  ### 2) Run inference
 
166
  - Inference framework (or API) and key settings (e.g., temperature, max tokens)
167
  - Any special post-processing applied (if applicable)
168
 
169
+ After validation, we will compute the official metrics on the hidden ground-truth and update the leaderboard.
 
video_understanding/README_cn.md CHANGED
@@ -20,7 +20,7 @@ FCMBench-Video v1.0 包含 **11,322** 组视频-问题对,覆盖 7 个任务
20
  ```
21
  FCMBench-Video_v1.0_Videos/
22
  └── video/
23
- ├── Construction/ # 中文原始视频(15 人 × 9 = 135 个片段)
24
  ├── Construction-en-US/ # 英文原始视频(30 人 × 9 = 265 个片段)
25
  ├── VPI-cn/ # 中文视觉提示注入视频(135 个片段)
26
  ├── VPI-cot-cn/ # 中文 VPI + 思维链视频(135 个片段)
@@ -34,7 +34,6 @@ FCMBench-Video_v1.0_Videos/
34
  |------|------|:------:|------|
35
  | `FCMBench-Video_v1.0_full.jsonl` | 全量 11,322 条(中英文合并) | ✗ | 公开发布,用于推理 |
36
  | `FCMBench-Video_v1.0_small.jsonl` | 分层 10% 采样(~1,135 条) | ✓ | 快速自测 / 调试诊断 |
37
- | `FCMBench-Video_v1.0_full-gt.jsonl` | 全量 11,322 条 | ✓ | 内部参考(不分发) |
38
 
39
  **任务分布**(7 类,覆盖感知与推理):
40
 
@@ -80,7 +79,7 @@ unzip FCMBench-Video_v1.0_Videos.zip
80
  ```
81
  FCMBench-Video_v1.0_Videos/
82
  └── video/
83
- ├── Construction/
84
  ├── Construction-en-US/
85
  ├── VPI-cn/
86
  ├── VPI-cot-cn/
@@ -89,7 +88,7 @@ FCMBench-Video_v1.0_Videos/
89
  ```
90
 
91
  指令文件的每条记录使用 `video_prefix` + `video_path` 来定位视频(例如:
92
- `"video_prefix": "Construction", "video_path": "yangyimiao/yangyimiao_20s_1.mp4"`),
93
  因此将 `--video_root` 指向 `FCMBench-Video_v1.0_Videos/` 目录即可。
94
 
95
  ### 2) 运行推理
@@ -162,4 +161,4 @@ python benchmark_eval.py --result_dir ./results
162
  - 推理框架(或 API)及关键参数设置(如 temperature、max tokens 等)
163
  - 是否使用了特殊的后处理(如适用)
164
 
165
- 我们验证通过后,会在隐藏的真值文件(`FCMBench-Video_v1.0_full-gt.jsonl`)上计算官方指标并更新排行榜。
 
20
  ```
21
  FCMBench-Video_v1.0_Videos/
22
  └── video/
23
+ ├── Construction-cn/ # 中文原始视频(15 人 × 9 = 135 个片段)
24
  ├── Construction-en-US/ # 英文原始视频(30 人 × 9 = 265 个片段)
25
  ├── VPI-cn/ # 中文视觉提示注入视频(135 个片段)
26
  ├── VPI-cot-cn/ # 中文 VPI + 思维链视频(135 个片段)
 
34
  |------|------|:------:|------|
35
  | `FCMBench-Video_v1.0_full.jsonl` | 全量 11,322 条(中英文合并) | ✗ | 公开发布,用于推理 |
36
  | `FCMBench-Video_v1.0_small.jsonl` | 分层 10% 采样(~1,135 条) | ✓ | 快速自测 / 调试诊断 |
 
37
 
38
  **任务分布**(7 类,覆盖感知与推理):
39
 
 
79
  ```
80
  FCMBench-Video_v1.0_Videos/
81
  └── video/
82
+ ├── Construction-cn/
83
  ├── Construction-en-US/
84
  ├── VPI-cn/
85
  ├── VPI-cot-cn/
 
88
  ```
89
 
90
  指令文件的每条记录使用 `video_prefix` + `video_path` 来定位视频(例如:
91
+ `"video_prefix": "Construction-cn", "video_path": "yangyimiao/yangyimiao_20s_1.mp4"`),
92
  因此将 `--video_root` 指向 `FCMBench-Video_v1.0_Videos/` 目录即可。
93
 
94
  ### 2) 运行推理
 
161
  - 推理框架(或 API)及关键参数设置(如 temperature、max tokens 等)
162
  - 是否使用了特殊的后处理(如适用)
163
 
164
+ 我们验证通过后,会在隐藏的真值文件上计算官方指标并更新排行榜。
video_understanding/instructions/FCMBench-Video_v1.0_full.jsonl CHANGED
The diff for this file is too large to render. See raw diff
 
video_understanding/instructions/FCMBench-Video_v1.0_small.jsonl CHANGED
The diff for this file is too large to render. See raw diff