PAL grounded subgoal training dataset

This repository contains 94,372 multimodal robot subgoal training samples from 16 tasks and all 768,897 PNG frames from the 1,600 source episodes. The JSONL directly references 240,855 of these images; the other 528,042 frames are also included so self-labeling can construct windows and four-step null-fill queries at timesteps absent from the original JSONL. Images are distributed as 32 lossless tar.zst archives: 16 original referenced-image archives and 16 additional full-timeline archives, grouped by task. Extract both sets before self-labeling.

Download and restore images

Install the Hugging Face CLI (pip install -U huggingface_hub) and the zstd command (brew install zstd on macOS or sudo apt-get install zstd on Debian/Ubuntu). Then, from an empty directory:

hf download qinghangliu/PAL --repo-type dataset --local-dir .
python unpack_images.py --root .

The helper requires Python 3.9 or later. It verifies archive SHA-256 checksums, validates every member path, and verifies all restored PNG files against the image manifest. It resumes safely when matching images already exist and refuses to overwrite different files. Four tasks are processed concurrently by default; use --workers 1 to reduce disk load. To verify archives without writing the extracted images:

python unpack_images.py --root . --verify-only

Image archives total 40,273,040,236 bytes (40.27 GB). Allow another 78.81 GB for the extracted PNG files (decimal GB), plus the JSONL and manifests. The extracted training JSONL plus all PNG files totals 79,017,133,037 bytes.

Directory layout

README.md
unpack_images.py
archives_manifest.json
archives/
  memer_BinFill.tar.zst
  ... (16 original task archives)
  full_frames/
    memer_BinFill.tar.zst
    ... (16 additional-frame task archives)
data/vlm_keyframe/
  memer_all/grounded_subgoal_train.jsonl
  image_manifest.jsonl
  dataset_statistics.json
  memer_BinFill/images/*.png       # restored by unpack_images.py
  ... (16 task image directories)

All image paths are relative to the dataset snapshot or PAL project root, for example data/vlm_keyframe/memer_BinFill/images/BinFill_ep0_step0.png. They are not relative to the JSONL's parent directory. Run training from the snapshot/project root, or join each image path to that root explicitly. The data/ directory can be copied into a PAL checkout without retaining machine-specific absolute paths.

Data format and counts

The UTF-8 JSONL has one sample per line with messages (system/user/assistant conversation), objects (grounding metadata), videos, and images. The four images entries preserve the original input-slot order. Sample order, messages, objects, supervision, and all non-image fields are unchanged from the source dataset; only image paths were rewritten.

Item	Count / bytes
Training samples	94,372
Tasks	16
Image references	377,488
Unique images directly referenced by JSONL	240,855
Additional timeline images	528,042
Total exported PNG files	768,897
Episodes with complete frame timelines	1,600
Extracted PNG bytes	78,805,233,789
Training JSONL bytes	211,899,248
Compressed image archive bytes	40,273,040,236

dataset_statistics.json includes per-task sample counts and examples. image_manifest.jsonl records the relative path, byte count, SHA-256, and containing archive of every exported image. full_frame_export_report.json records frame continuity checks; self_label_frame_coverage.json records checks using the actual self-label window functions. archives_manifest.json records each archive's checksum, byte count, and image count. The PNG bytes are preserved exactly; images are not re-encoded or resized.

Why image references repeat

There are 136,633 references beyond the 240,855 unique image paths:

• Within a sample: four input slots sometimes point to the same frame, especially at the beginning of an episode. There are 19,254 extra references within samples, affecting 6,922 samples. For example, the opening BinFill samples use BinFill_ep0_step0.png in all four slots.
• Across samples: adjacent samples share history/window frames. This accounts for another 117,379 references; 54,610 images occur in more than one sample. For example, VideoRepick_ep4_step255.png appears 55 times across 31 samples.

Each unique image path is stored once. Repeated entries inside the JSONL are intentionally retained because their position has input-slot meaning. Distinct paths are not merged based on visual similarity or identical content.

Provenance and integrity

The source is grounded_subgoal_train.jsonl from the PAL preprocessed training data. The portable copy includes every PNG frame of the episodes referenced by that file, including conditioning and execution frames that are not direct JSONL references. All 768,897 copied images were checked against their source using SHA-256. All episode and per-phase timelines are contiguous; the JSONL contents have not changed.

Training JSONL SHA-256 after path rewriting:

e8885aed1962f37c0f941d939456572125a982c4066607353bde12af627c025e

This dataset repository is separate from the model repository qinghangliu/PAL.

Correction to the initial export

The initial export included only JSONL-referenced images. That was sufficient for reading the original supervised samples but incomplete for self-labeling, which scans sibling images and queries additional timesteps. For example, InsertPeg_ep14_condition_step8.png existed in the source but was omitted by that filter. The corrected export preserves every source PNG frame.

If you already downloaded and extracted the original 16 archives, run the download command again to obtain the updated helper/manifests and archives/full_frames/, then run python unpack_images.py --root .. Matching existing images are checked and preserved. The 16 additional archives total 27,386,051,700 bytes (27.39 GB); the original archives and training JSONL do not need to be downloaded again if cached locally.

Keeping the missing-frame guard in self-labeling is useful for other incomplete inputs, but skipping missing targets is no longer required for this dataset's four-step query grid.