Update model card
Browse files
README.md
CHANGED
|
@@ -76,50 +76,10 @@ The tokenizer vocabulary has 151,671 entries. It preserves all 13 Qwen3-VL addit
|
|
| 76 |
|
| 77 |
The TimePLE codec uses a 128×128 canonical span grid, a duration-adaptive residual decoder, and a learned MLP interface adapter to the 4096-dimensional language-model representation.
|
| 78 |
|
| 79 |
-
## Training
|
| 80 |
-
|
| 81 |
-
The released checkpoint is the validation-selected endpoint of the second full-parameter supervised fine-tuning stage. Stage 2 continued from the selected stage-1 checkpoint and unfroze the TimePLE encoder and decoder.
|
| 82 |
-
|
| 83 |
-
Key stage-2 settings:
|
| 84 |
-
|
| 85 |
-
| Setting | Value |
|
| 86 |
-
|---|---:|
|
| 87 |
-
| Training examples | 92,215 |
|
| 88 |
-
| Validation examples | 3,000 |
|
| 89 |
-
| Epochs | 1 |
|
| 90 |
-
| Global steps | 145 |
|
| 91 |
-
| Learning rate | 3e-6 |
|
| 92 |
-
| Scheduler | Cosine |
|
| 93 |
-
| Per-device batch size | 2 |
|
| 94 |
-
| Gradient accumulation | 20 |
|
| 95 |
-
| Distributed world size | 16 |
|
| 96 |
-
| Precision | bfloat16 |
|
| 97 |
-
| Optimizer | AdamW (`beta1=0.9`, `beta2=0.95`, weight decay `0.1`) |
|
| 98 |
-
| Distributed strategy | DeepSpeed ZeRO-2 |
|
| 99 |
-
| Random/data seeds | 42 / 42 |
|
| 100 |
-
|
| 101 |
-
Training used language-model loss together with TimePLE distribution-focal, IoU, and duration-aware boundary objectives. Videos were sampled at 6 FPS with 4–400 frames and bounded to 32–100 video tokens by the training data pipeline.
|
| 102 |
-
|
| 103 |
-
The released training annotations and benchmark annotations are available in [KlingTeam/TimePLE-Dataset](https://huggingface.co/datasets/KlingTeam/TimePLE-Dataset). Licensed source videos are not redistributed.
|
| 104 |
-
|
| 105 |
-
## Validation used for checkpoint selection
|
| 106 |
-
|
| 107 |
-
Checkpoint 145 was selected by the highest temporal-span IoU on the stage-2 validation mixture. These figures are training-time validation metrics, not a claim of performance on an official test split.
|
| 108 |
-
|
| 109 |
-
| Metric | Value |
|
| 110 |
-
|---|---:|
|
| 111 |
-
| Span IoU | 0.5210076190 |
|
| 112 |
-
| Mean absolute boundary error, seconds | 13.9402037276 |
|
| 113 |
-
| Evaluation loss | 0.0877585486 |
|
| 114 |
-
|
| 115 |
-
The original training logs used the legacy metric prefix `cis_span_duration_adaptive`; the released architecture and files use the final `timeple` names.
|
| 116 |
-
|
| 117 |
## Intended use
|
| 118 |
|
| 119 |
This model is intended for research and evaluation of temporal localization in videos from a textual event query. Suitable uses include moment retrieval experiments, temporal-grounding ablations, and further task-specific fine-tuning.
|
| 120 |
|
| 121 |
-
It is not designed for safety-critical decisions, surveillance decisions about people, or reliable frame-exact annotation without human review.
|
| 122 |
-
|
| 123 |
## Limitations
|
| 124 |
|
| 125 |
- Temporal quality depends on frame sampling, video duration metadata, and the inference pipeline's video preprocessing.
|
|
|
|
| 76 |
|
| 77 |
The TimePLE codec uses a 128×128 canonical span grid, a duration-adaptive residual decoder, and a learned MLP interface adapter to the 4096-dimensional language-model representation.
|
| 78 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79 |
## Intended use
|
| 80 |
|
| 81 |
This model is intended for research and evaluation of temporal localization in videos from a textual event query. Suitable uses include moment retrieval experiments, temporal-grounding ablations, and further task-specific fine-tuning.
|
| 82 |
|
|
|
|
|
|
|
| 83 |
## Limitations
|
| 84 |
|
| 85 |
- Temporal quality depends on frame sampling, video duration metadata, and the inference pipeline's video preprocessing.
|