GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
georgeliu59
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
georgeliu59
Model Tree
Qwen/Qwen3-VL-8B-Instruct
Input Modalities
Output Modalities
Supported Functionality
More information needed
More information needed
More information needed
The following hyperparameters were used during training:
The final five-epoch model was evaluated on the preserved October test split: 9,522 examples (4,761 Yes and 4,761 No), 1,069 videos, and 244 tasks, with no October train/test video overlap. The data producer caps held-out examples per task; this is not a random 20% of the full expanded dataset.
Answer accuracy uses the first freely predicted token. Ranking scores use full-vocabulary Yes probability; pairs follow the original exported pairing and ties are errors. Sampling matches training: 8 fps, at most 128 frames, 65,536 pixels per frame, qwen3_vl_nothink, BF16, and SDPA. All examples retained their original order and fit below the 8,192-token cutoff. Distributed sampler padding was verified and excluded.
Historical comparison uses 7,020 current-label examples across 797 videos present in the retained February test file and absent from its rare-label training file (2,593 complete pairs). The full October test includes 255 videos used by the historical rare-label stage; use the shared subset for that comparison. The original complete February training JSON is unavailable, so historical held-out membership relies on the retained test file and full-test historical comparisons may include training exposure.
| Model | Shared answer accuracy | Shared mAP | Shared pairwise accuracy |
|---|---|---|---|
| october | 81.18% | 92.44% | 92.23% |
| february | 80.54% | 91.42% | 90.84% |
| february_rare | 80.21% | 90.67% | 88.81% |
The October-minus-February shared-test accuracy difference has a 95% paired video-bootstrap interval of -0.40 to 1.69 percentage points, so its small accuracy advantage is not conclusive.
Evaluated model revision: fc27ebd93dd703070e734a4f36520a0d916cc9cd. See evaluation_comparison.json for exact model revisions, metrics, and paired video-bootstrap confidence intervals. Only aggregate evaluation files and this model card were updated; weights and training assets are unchanged.