Update README.md
Browse files
README.md
CHANGED
|
@@ -11,7 +11,7 @@ tags:
|
|
| 11 |
- llama.cpp
|
| 12 |
---
|
| 13 |
|
| 14 |
-
# **TimeLens-8B**
|
| 15 |
|
| 16 |
> TimeLens-8B from TencentARC is an 8B-parameter multimodal vision-language model fine-tuned from Qwen3-VL-8B-Instruct using a novel RLVR (reinforcement learning with verifiable rewards) recipe on the high-quality TimeLens-100K VTG dataset, achieving state-of-the-art video temporal grounding performance among open-source models with 72.0% R1@0.3 (Charades-TimeLens), 64.5% R1@0.3 (ActivityNet-TimeLens), and 75.6% R1@0.3 (QVHighlights-TimeLens), significantly outperforming baselines like Qwen3-VL-8B-Instruct and Qwen2.5-VL-7B. Designed for precise localization of visual events described by natural language queries, it outputs timestamped segments in the format "The event happens in <start time> - <end time> seconds" using low FPS=2 sampling (min_pixels=642828, total_pixels=143362828) for efficient video processing via Transformers with Flash-Attention-2 support. Released with code, project page, and TimeLens-Bench evaluation suite, it excels on Charades-TimeLens, ActivityNet-TimeLens, and QVHighlights-TimeLens leaderboards for research in video understanding, temporal reasoning, and event detection.
|
| 17 |
|
|
|
|
| 11 |
- llama.cpp
|
| 12 |
---
|
| 13 |
|
| 14 |
+
# **TimeLens-8B-GGUF**
|
| 15 |
|
| 16 |
> TimeLens-8B from TencentARC is an 8B-parameter multimodal vision-language model fine-tuned from Qwen3-VL-8B-Instruct using a novel RLVR (reinforcement learning with verifiable rewards) recipe on the high-quality TimeLens-100K VTG dataset, achieving state-of-the-art video temporal grounding performance among open-source models with 72.0% R1@0.3 (Charades-TimeLens), 64.5% R1@0.3 (ActivityNet-TimeLens), and 75.6% R1@0.3 (QVHighlights-TimeLens), significantly outperforming baselines like Qwen3-VL-8B-Instruct and Qwen2.5-VL-7B. Designed for precise localization of visual events described by natural language queries, it outputs timestamped segments in the format "The event happens in <start time> - <end time> seconds" using low FPS=2 sampling (min_pixels=642828, total_pixels=143362828) for efficient video processing via Transformers with Flash-Attention-2 support. Released with code, project page, and TimeLens-Bench evaluation suite, it excels on Charades-TimeLens, ActivityNet-TimeLens, and QVHighlights-TimeLens leaderboards for research in video understanding, temporal reasoning, and event detection.
|
| 17 |
|