gsarch/foveated-vlm-chart-sft

Chart fixation SFT and teacher-forced replay

Qwen3-VL chart trajectories with a 3× overview and native-resolution crops.

• sft.json.gz: 1,969 training trajectories, with 41,933 crop actions.
• teacher.json.gz: 1,969 historical replay trajectories with upscaled sources.
• sources/: 1,945 shared lossless source images.
• manifest.json: source hashes and reconstruction checks.

Each record contains the question, assistant point requests, final answer, image geometry, and hashes of the original overview and crop JPEGs. Coordinates in point_2d are normalized to [0, 1000]. Each crop covers about 2% of its source.

Use scripts/prepare_chart.py from the accompanying foveated_vlm code to download this bundle and regenerate the saved images. The script checks every JPEG against its original hash. Use Pillow 11.3.0 for reconstruction.

The SFT and teacher-forced records describe different image preparations; keep them separate. Teacher forcing provides the recorded history and scores the generated answer against gt_answer. These are supplied trajectories, not a held-out ChartQA test split.

The source chart images retain the terms of their original datasets.