AnchorSR-Qwen3.5-9B-SFT-v4
Version: v4. Full-parameter SFT on TrainingData_v4, starting from Qwen/Qwen3.5-9B
(revision c202236235762e1c871ad0ccb60c8ee5ba337b9a). This is the final checkpoint after 2 epochs, 1,000 optimizer steps,
with a global batch size of 64.
The training set has 32,000 examples per epoch, with epoch-rotated 16/24/32-frame video views. Validation uses the canonical 8,000-example V4 split. Training uses BF16, a learning rate of 1e-6, cosine scheduling with 3% warmup, an 8,192-token cutoff, and a numeric token loss share of 0.2. The final validation loss is 0.14332369 under this training objective; this is not a spatial-reasoning benchmark score.
The repository contains the final full model weights, tokenizer, chat template, and processor
configuration exported with Transformers 5.8.0. training_metadata.json records the settings
and final metrics.
from huggingface_hub import snapshot_download
model_path = snapshot_download("AnchorSR/AnchorSR-Qwen3.5-9B-SFT-v4")
Training completed at 2026-09-17T06:37:44.773302+00:00 (UTC). Online training log.
Validation loss by epoch
| Epoch | Optimizer step | Validation loss |
|---|---|---|
| 1 | 500 | 0.14716931 |
| 2 | 1000 | 0.14332369 |