HughieHu/drpixel-train

DR Pixel training data

This is the training set behind the trained DR Pixel model in DR Pixel: Page-Native Deep Research with Vision-Language Models: 2,496 trajectories of GPT-5.4 acting as the DR Pixel agent, with the per-step loss weights from counterfactual credit reweighting, exactly as used to fine-tune Qwen3.6-27B.

The model trained on it is HughieHu/DR-Pixel-Trained. The training code is in github.com/HughieHu/DRPixel (train/), and the benchmark and retrieval corpus are in HughieHu/drpixel-bench.

Layout

train.jsonl                          28 MB   2,496 trajectories
image-shards/part-000 … part-003.tar  3.2 GB  the 17,739 page images they reference
credit/trajectories.jsonl            28 MB   the same trajectories before weighting, with raw credits
credit/rank_pool.json               0.5 MB   the ranking pools that turn those credits into weights
SHA256SUMS                                   checksums of the files above

How the trajectories were made

Each trajectory is a run of GPT-5.4 as the DR Pixel agent on a training question. The training questions come from the same pipeline as the benchmark, taken before human annotation, over a disjoint pool of papers: no paper that supports a training question supports a benchmark question. Only runs that reached a correct final answer were kept.

Format

One chat-format trajectory per line:

{"messages": [{"role": "system", "content": "..."},
              {"role": "user", "content": "QUESTION: ..."},
              {"role": "assistant", "content": "<think>...</think>", "loss_scale": 0.84},
              {"role": "assistant", "content": "<call_tool name=\"search_page\">...</call_tool>", "loss_scale": 1.31},
              {"role": "tool", "content": "<tool_output><page id=P0_0 doc=... page=4><image>...</tool_output>"},
              ...,
              {"role": "assistant", "content": "<think>...</think><answer>...</answer>", "loss_scale": 1.5}],
 "images": ["images/<hash>.jpg", ...]}

Retrieved pages arrive as tool turns, with each page image marked <image> in the order of images. Each assistant turn of a search step is split into its reasoning and its action, and each part carries its own loss_scale. The counterfactual credit of the step is converted to a rank-based weight, where 0 masks the step and the largest weight is 2. A reasoning step is kept only if its action is kept and it raises the likelihood of that action at least as much as the median reasoning step does. The final answer has weight 1.5.

Credits

credit/trajectories.jsonl holds the same 2,496 trajectories before weighting, in the same order and with the same text and image paths as train.jsonl. Search steps carry their raw counterfactual credits, cf_query for the search and cf_think for its reasoning, and na_think, the gain in the action's likelihood from the reasoning over an empty reasoning block. All three were computed under the frozen Qwen3.6-27B. The scripts in the training code rebuild the training set from this file without a GPU:

python make_weights.py --shards credit/trajectories.jsonl --rank_pool credit/rank_pool.json --out weighted.jsonl
python split_think.py  weighted.jsonl split.jsonl
python build_train.py  --split split.jsonl --out train.jsonl

Images

Extract the shards inside the dataset directory; this creates images/, and the paths in train.jsonl are relative to that directory:

for t in image-shards/*.tar; do tar -xf "$t"; done

The page images are stored as they were used in training. The training recipe caps every image at 602,112 pixels. In long trajectories, the images of earlier steps are stored at a half, a quarter, an eighth or a sixteenth of that, so that the whole trajectory fits the training length.

Training

train/README.md in the code repository shows how to train on this data and how the weights were computed.

License

The trajectories are released for research under CC BY 4.0, subject to the terms of the model that generated them. The page images are reproduced from the source papers for research use, and copyright stays with their authors.

Citation

@misc{hu2026drpixel,
  title  = {DR Pixel: Page-Native Deep Research with Vision-Language Models},
  author = {Hu, Tiansheng and Wang, Yichuan and Talwar, James V. and Tomlin, Nicholas and Zhao, Chen and Min, Sewon},
  year   = {2026}
}