Video-Text-to-Text
Transformers
Safetensors
English
Chinese
moss_vl
feature-extraction
SFT
Video-Understanding
Image-Understanding
MOSS-VL
OpenMOSS
multimodal
video
vision-language
custom_code
Instructions to use OpenMOSS-Team/MOSS-VL-Instruct-0408 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use OpenMOSS-Team/MOSS-VL-Instruct-0408 with Transformers:
# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("OpenMOSS-Team/MOSS-VL-Instruct-0408", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
Upload folder using huggingface_hub
Browse files
README.md
CHANGED
|
@@ -54,7 +54,6 @@ Built on top of MOSS-VL-Base-0408 through supervised fine-tuning (SFT), this che
|
|
| 54 |
<img src="assets/structure.png" alt="MOSS-VL Architecture" width="90%"/>
|
| 55 |
</p>
|
| 56 |
|
| 57 |
-
|
| 58 |
## 🧩 Absolute Timestamps
|
| 59 |
|
| 60 |
To ensure the model accurately perceives the pacing and duration of events, **MOSS-VL-Instruct-0408** injects **absolute timestamps** alongside each sampled frame, grounding the reasoning process in a **precise temporal reference**.
|
|
@@ -71,7 +70,6 @@ MOSS-VL utilizes Cross-attention Rotary Position Embedding (XRoPE), tailored to
|
|
| 71 |
<img src="assets/3d-rope.png" alt="MOSS-VL mRoPE Architecture Illustration" width="80%"/>
|
| 72 |
</p>
|
| 73 |
|
| 74 |
-
|
| 75 |
## 📊 Model Performance
|
| 76 |
|
| 77 |
We conducted a comprehensive evaluation of **MOSS-VL-Instruct-0408** across four key dimensions: Multimodal Perception, Document/OCR, Multimodal Reasoning, and Video Understanding. The results demonstrate that MOSS-VL achieves outstanding performance, particularly excelling in **general multimodal perception** and **complex video analysis**.
|
|
@@ -90,10 +88,7 @@ We conducted a comprehensive evaluation of **MOSS-VL-Instruct-0408** across four
|
|
| 90 |
<img src="assets/MOSS-VL-benchmark.png" alt="MOSS-VL Benchmark Results" width="100%"/>
|
| 91 |
</p>
|
| 92 |
|
| 93 |
-
|
| 94 |
## 🚀 Quickstart
|
| 95 |
-
|
| 96 |
-
|
| 97 |
### 🛠️ Requirements
|
| 98 |
|
| 99 |
Installation commands:
|
|
@@ -294,9 +289,6 @@ texts = [item["text"] for item in result["results"]]
|
|
| 294 |
|
| 295 |
</details>
|
| 296 |
|
| 297 |
-
|
| 298 |
-
|
| 299 |
-
|
| 300 |
## 🚧 Limitations and Future Work
|
| 301 |
|
| 302 |
MOSS-VL-Instruct-0408 represents an early milestone in the MOSS-VL roadmap, and we're actively working on several directions to push it further:
|
|
|
|
| 54 |
<img src="assets/structure.png" alt="MOSS-VL Architecture" width="90%"/>
|
| 55 |
</p>
|
| 56 |
|
|
|
|
| 57 |
## 🧩 Absolute Timestamps
|
| 58 |
|
| 59 |
To ensure the model accurately perceives the pacing and duration of events, **MOSS-VL-Instruct-0408** injects **absolute timestamps** alongside each sampled frame, grounding the reasoning process in a **precise temporal reference**.
|
|
|
|
| 70 |
<img src="assets/3d-rope.png" alt="MOSS-VL mRoPE Architecture Illustration" width="80%"/>
|
| 71 |
</p>
|
| 72 |
|
|
|
|
| 73 |
## 📊 Model Performance
|
| 74 |
|
| 75 |
We conducted a comprehensive evaluation of **MOSS-VL-Instruct-0408** across four key dimensions: Multimodal Perception, Document/OCR, Multimodal Reasoning, and Video Understanding. The results demonstrate that MOSS-VL achieves outstanding performance, particularly excelling in **general multimodal perception** and **complex video analysis**.
|
|
|
|
| 88 |
<img src="assets/MOSS-VL-benchmark.png" alt="MOSS-VL Benchmark Results" width="100%"/>
|
| 89 |
</p>
|
| 90 |
|
|
|
|
| 91 |
## 🚀 Quickstart
|
|
|
|
|
|
|
| 92 |
### 🛠️ Requirements
|
| 93 |
|
| 94 |
Installation commands:
|
|
|
|
| 289 |
|
| 290 |
</details>
|
| 291 |
|
|
|
|
|
|
|
|
|
|
| 292 |
## 🚧 Limitations and Future Work
|
| 293 |
|
| 294 |
MOSS-VL-Instruct-0408 represents an early milestone in the MOSS-VL roadmap, and we're actively working on several directions to push it further:
|