lat_res

# LVR current training data This package contains the annotation JSON files used by the current Latent Visual Reasoning pipeline. It contains no model checkpoints, raw images, oracle-evidence tensors, rollouts, credentials, or sealed test data. ## Files | Stage | File | Rows | Status | |---|---|---:|---| | Stage-A SFT | `data/sft/train_stage_a_exact_used.json` | 35,232 | Exact historical annotations; includes 53 non-strict bbox entries | | Stage-A SFT | `data/sft/train_stage_a_strict_recommended.json` | 35,179 | Recommended strict positive-area version | | Localization mid-training | `data/mid_training/treevgr_rl_full_pointer_midtrain.json` | 29,310 | Exact pointer mid-training annotations | | Paired RL | `data/rl/zwz_train20k_seed731_exact_used.json` | 20,000 | Exact Oracle/Prediction RL training subset | | RL evaluation | `data/eval/zwz_val_decodeclean1999_seed731.json` | 1,999 | Strict decoded-RGB-clean validation; not training data | Every row uses the LVR contract: ```json { "images": ["relative/path.jpg"], "bboxes": [[0.1, 0.2, 0.4, 0.6]], "conversations": [ {"from": "human", "value": "<image>\nQuestion"}, {"from": "gpt", "value": "<think><lvr></think><answer>Gold</answer>"} ] } ``` Coordinates are normalized `xyxy`. One `<lvr>` corresponds to one bbox; the runtime expands each `<lvr>` into four continuous `<vis>` positions. ## Images are intentionally excluded Resolve each relative path against the corresponding upstream image root: - SFT: download `lmms-lab/LLaVA-NeXT-Data`; paths such as `images/gqa/...` are relative to that dataset root. - Mid-training: download `HaochenWang/TreeVGR-RL-37K` and extract its `images.tar.gz`; paths are `images/<id>.jpg`. - RL/eval: download `inclusionAI/ZwZ-RL-VQA` at revision `fec404cb7bd5a18458e5add2802a5fb183981d14` and extract the clean `original_images` shards. Never substitute the red-box overlay images. The JSON paths were rewritten from private absolute filesystem paths to these portable relative paths. `MANIFEST.json` records both the exact source SHA256 used in training and the packaged SHA256. ## Data discipline - The exact historical Stage-A file is retained for reproducibility, but paper-final retraining should use the strict 35,179-row file. - The historical ZwZ train20k has 17 beyond-first decoded-RGB duplicates (0.085%). This is disclosed rather than silently changing the data used by the run. - The 1,999-row validation set has one row per decoded RGB group and zero decoded-RGB overlap with train20k. - No sealed final split is included. ## Sources and licenses The transformed annotations follow the Apache-2.0 dataset cards of: - [HaochenWang/TreeVGR-SFT-35K](https://huggingface.co/datasets/HaochenWang/TreeVGR-SFT-35K) - [HaochenWang/TreeVGR-RL-37K](https://huggingface.co/datasets/HaochenWang/TreeVGR-RL-37K) - [inclusionAI/ZwZ-RL-VQA](https://huggingface.co/datasets/inclusionAI/ZwZ-RL-VQA) Images are not redistributed here and may retain their original source-specific terms. Cite TreeVGR (arXiv:2507.07999) and Zooming without Zooming (arXiv:2602.11858) when using the corresponding annotations.