Evaluation
Evaluation uses the thinking prompting protocol. Scores are percentages.
Table with columns: Benchmark, Score| Benchmark | Score |
|---|
| BLINK (micro) | 74.75 |
| MUIR (micro) | 80.08 |
| VLM2 (8-task macro) | 81.38 |
| Mean of the three benchmarks | 78.74 |
These results describe this checkpoint under the stated protocol; direct-answer evaluations use a different protocol and are not interchangeable.
Files and use
The repository includes safetensors weights, their index, tokenizer files, chat template, and visual preprocessing configuration. Use the inference implementation for the base model with this repository's processor and chat template. The evaluated response format contains reasoning in <thinking>...</thinking> and the selected option in <answer>...</answer>.
SHA256SUMS records local file checksums. Model shard hashes are checked against Hugging Face's uploaded LFS metadata after upload. No training images, source data, account credentials, or optimizer states are included.
The base-model license and any included license file apply to this derivative checkpoint.