AnchorSR-Qwen3.5-9B-SFT

AnchorSR-Qwen3.5-9B-SFT

Full-parameter supervised fine-tuning of Qwen3.5-9B on AnchorSR v3.

Training provenance

• Run: qwen3_5_9b_sft_20260908_120030_673008
• Training examples: 32,000; validation examples: 8,000.
• Completed epochs: 2.
• Optimizer updates: 2000 / 2000.
• Actual global batch size for this completed run: 32.
• Precision / distributed optimizer: BF16 / DeepSpeed ZeRO-2.
• Video sampling: uniform 16 frames.
• Numeric-answer loss share: 0.2.
• Final validation loss: 0.13139974.

The validation loss is a training metric, not a task benchmark score; losses across different model tokenizers are not directly comparable.

Contents

Final model weights, tokenizer, processor and generation configuration are included. Intermediate checkpoints and optimizer states are not included. This export supports model loading and inference; it is not an exact optimizer resume checkpoint. Preserve the supplied chat template and multimodal processor.

This model is intended for research on spatial reasoning. Downstream benchmark accuracy and general deployment behavior have not been established by these training-completion checks.