Dataset Preview

Split (1)

train

The full dataset viewer is not available (click to read why). Only showing a preview of the rows.

The dataset generation failed

Error code:   DatasetGenerationError
Exception:    TypeError
Message:      int() argument must be a string, a bytes-like object or a real number, not 'NoneType'
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1531, in _prepare_split_single
                  for key, record in generator:
                                     ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 127, in _generate_examples
                  for example_idx, example in enumerate(self._get_pipeline_from_tar(tar_path, tar_iterator)):
                                              ~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 32, in _get_pipeline_from_tar
                  for filename, f in tar_iterator:
                                     ^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/track.py", line 49, in __iter__
                  for x in self.generator(*self.args):
                           ~~~~~~~~~~~~~~^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/file_utils.py", line 1400, in _iter_from_urlpath
                  with xopen(urlpath, "rb", download_config=download_config, block_size=0) as f:
                       ~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/file_utils.py", line 977, in xopen
                  file_obj = fs.open(paths[0], mode)
                File "<string>", line 3, in open
                File "/usr/local/lib/python3.14/unittest/mock.py", line 1176, in __call__
                  return self._mock_call(*args, **kwargs)
                         ~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/unittest/mock.py", line 1180, in _mock_call
                  return self._execute_mock_call(*args, **kwargs)
                         ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/unittest/mock.py", line 1247, in _execute_mock_call
                  result = effect(*args, **kwargs)
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 786, in wrapped
                  tracker.files[urlpath] = {"read": 0, "size": int(f.size)}
                                                               ~~~^^^^^^^^
              TypeError: int() argument must be a string, a bytes-like object or a real number, not 'NoneType'
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1393, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1571, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

jsonl unknown	__key__ string	__url__ string
"eyJpZCI6ICJwbC1ob3VzZS0wMDEtcjQtYWlyd2FsbC0wLXIwIiwgIm9iamVjdF9jaG9pY2VzIjogMiwgIm1pbl9iZXN0X3B4Ijo(...TRUNCATED)	traj_v4_clean/route_quality	"hf://datasets/kimperyang/vlmauditor-data-v4@05a3c91b7758746c863f971047e1eb8aead93db2/clean/traj_v4_(...TRUNCATED)
"eyJpZCI6ICJwbC1zY2hvb2wtMDExLXI0LW1vdmVkLTAtcjAiLCAic291cmNlIjogInI0IiwgImZvcm1hdCI6ICJ2NCIsICJzcGx(...TRUNCATED)	traj_v4_clean/test_clean	"hf://datasets/kimperyang/vlmauditor-data-v4@05a3c91b7758746c863f971047e1eb8aead93db2/clean/traj_v4_(...TRUNCATED)
"eyJpZCI6ICJwbC1zY2hvb2wtMDExLXI0LW1vdmVkLTAtcjAiLCAic291cmNlIjogInI0IiwgImZvcm1hdCI6ICJ2NCIsICJzcGx(...TRUNCATED)	traj_v4_clean/test_cot_clean	"hf://datasets/kimperyang/vlmauditor-data-v4@05a3c91b7758746c863f971047e1eb8aead93db2/clean/traj_v4_(...TRUNCATED)
"eyJpZCI6ICJwbC1vZmZpY2UtMDAyLXI0LXdhbGxjbGlwLTAtcjExIiwgInNvdXJjZSI6ICJyNCIsICJmb3JtYXQiOiAidjQiLCA(...TRUNCATED)	traj_v4_clean/train_clean	"hf://datasets/kimperyang/vlmauditor-data-v4@05a3c91b7758746c863f971047e1eb8aead93db2/clean/traj_v4_(...TRUNCATED)
"eyJpZCI6ICJwbC1vZmZpY2UtMDAyLXI0LXdhbGxjbGlwLTAtcjExIiwgInNvdXJjZSI6ICJyNCIsICJmb3JtYXQiOiAidjQiLCA(...TRUNCATED)	traj_v4_clean/train_cot_clean	"hf://datasets/kimperyang/vlmauditor-data-v4@05a3c91b7758746c863f971047e1eb8aead93db2/clean/traj_v4_(...TRUNCATED)
"eyJpZCI6ICJwbC1tYXJrZXQtMDEwLXI0LW5vc29saWQtMy1yMTUiLCAic291cmNlIjogInI0IiwgImZvcm1hdCI6ICJ2NCIsICJ(...TRUNCATED)	traj_v4_clean/val_clean	"hf://datasets/kimperyang/vlmauditor-data-v4@05a3c91b7758746c863f971047e1eb8aead93db2/clean/traj_v4_(...TRUNCATED)
"eyJpZCI6ICJwbC1tYXJrZXQtMDEwLXI0LW5vc29saWQtMy1yMTUiLCAic291cmNlIjogInI0IiwgImZvcm1hdCI6ICJ2NCIsICJ(...TRUNCATED)	traj_v4_clean/val_cot_clean	"hf://datasets/kimperyang/vlmauditor-data-v4@05a3c91b7758746c863f971047e1eb8aead93db2/clean/traj_v4_(...TRUNCATED)
null	traj_v4_clean/stats	"hf://datasets/kimperyang/vlmauditor-data-v4@05a3c91b7758746c863f971047e1eb8aead93db2/clean/traj_v4_(...TRUNCATED)
"eyJpZCI6ICJwbC1zY2hvb2wtMDExLXI0LW1vdmVkLTAtcjAiLCAic291cmNlIjogInI0IiwgImZvcm1hdCI6ICJ2NCIsICJzcGx(...TRUNCATED)	data/traj_v4/test	"hf://datasets/kimperyang/vlmauditor-data-v4@05a3c91b7758746c863f971047e1eb8aead93db2/part1/data.tar(...TRUNCATED)
"eyJpZCI6ICJwbC1zY2hvb2wtMDExLXI0LW1vdmVkLTAtcjAiLCAic291cmNlIjogInI0IiwgImZvcm1hdCI6ICJ2NCIsICJzcGx(...TRUNCATED)	data/traj_v4/test_cot	"hf://datasets/kimperyang/vlmauditor-data-v4@05a3c91b7758746c863f971047e1eb8aead93db2/part1/data.tar(...TRUNCATED)

End of preview.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

VLMAuditor data v4 (transfer through Hugging Face)

Built on katrina1 on 2026-10-23 from out/data4-20261021/release (the final v4 build of 2026-10-21 12:06 UTC). The repo's README §5b describes the data, the trainer and what to watch.

Folder	Content	Size
`part1/data.tar`	every jsonl: `data/traj_v4/{train,val,test}.jsonl` (no thoughts), `data/traj_v4/{train,val,test}_cot.jsonl` (with thoughts), `data/ident_v4/{train,val,test}.jsonl`, plus `files_trainval.txt`, `files_test.txt`, `manifest.json`, `README_DATA.md`	1.5 GB
`part1/ident-images-000..007.tar`	every image that the ident train/val files use: 394,566 files	36.8 GB
`part2/traj-images-000..015.tar`	the rest of the images that the no-thought trajectory train/val files use: 768,048 files. The `_cot` files use a subset of the same images.	70.9 GB
`clean/traj_v4_clean.tar`	the trajectory files without routes in which the explorer went to an object it had barely seen (best view < 300 px): `traj_v4_clean/{train,val,test}_clean.jsonl`, `..._cot_clean.jsonl`, `route_quality.jsonl` (per-route numbers for other thresholds), `stats.json`. Train keeps 12,803 of 18,816 routes. See `docs/data-generation-20261023.md` §9 in the repo.	0.5 GB

Paths inside the image tars are repo-relative (out/..., examples/...). Extract every image tar into ONE folder, IMAGE_ROOT; the jsonl files then need no rewriting.
Not included: the test images (21.2 GB; only for offline test metrics), the Gemini-with-rules trajectories and the planner trajectories (80 actions; the inference path still assumes 40).

On the training machine

hf download kimperyang/vlmauditor-data-v4 --repo-type dataset --include "part1/*" --local-dir /data/hf-v4
cd /data/hf-v4/part1 && sha256sum -c SHA256SUMS
mkdir -p /data/v4 /data/v4-images && tar -xf data.tar -C /data/v4
for t in ident-images-*.tar; do tar -xf "$t" -C /data/v4-images; done
# from now on: DATA=/data/v4/data  IMAGE_ROOT=/data/v4-images

Part 2 the same way (--include "part2/*", then extract traj-images-*.tar into the same /data/v4-images). The filtered trajectory files: --include "clean/*", check clean/SHA256SUMS, then tar -xf clean/traj_v4_clean.tar -C /data/v4 (gives /data/v4/traj_v4_clean/).

Run 1: identification only (needs part 1 only)

From the VLMAuditor repo root. Empty TRAIN_TRAJ / VAL_TRAJ make train_v4.sh train on the ident files alone (EPOCHS then counts passes over ident).

envs/train/bin/python tools/nav/check_labels4.py --model Qwen/Qwen3.5-9B --image-root /data/v4-images --ident /data/v4/data/ident_v4/val.jsonl --n 20
TRAIN_TRAJ= VAL_TRAJ= DATA=/data/v4/data IMAGE_ROOT=/data/v4-images OUT=runs/smoke-ident vlmauditor/launch/train_v4.sh --max-steps 3 --save-every 3 --keep 1 --val-n 8 --smoke
TRAIN_TRAJ= VAL_TRAJ= DATA=/data/v4/data IMAGE_ROOT=/data/v4-images OUT=runs/qwen35-9b-ident vlmauditor/launch/train_v4.sh

Purpose: does our identification data fix the weakest part (recognising the bug in a frame)? Watch val_ident. It also checks the machine: environment, speed (sec) and memory (mem_gb) in train_log.jsonl.

Run 2: the full v4 recipe (needs parts 1 and 2)

No-thought trajectories plus ident, MIX 1:6 (the launcher defaults). Use the filtered trajectories (clean/); the unfiltered files are the comparison:

envs/train/bin/python tools/nav/check_labels4.py --model Qwen/Qwen3.5-9B --image-root /data/v4-images --traj /data/v4/data/traj_v4/val.jsonl --ident /data/v4/data/ident_v4/val.jsonl --n 20
TRAIN_TRAJ=/data/v4/traj_v4_clean/train_clean.jsonl VAL_TRAJ=/data/v4/traj_v4_clean/val_clean.jsonl DATA=/data/v4/data IMAGE_ROOT=/data/v4-images OUT=runs/smoke-v4 vlmauditor/launch/train_v4.sh --max-steps 3 --save-every 3 --keep 1 --val-n 8 --smoke
TRAIN_TRAJ=/data/v4/traj_v4_clean/train_clean.jsonl VAL_TRAJ=/data/v4/traj_v4_clean/val_clean.jsonl DATA=/data/v4/data IMAGE_ROOT=/data/v4-images OUT=runs/qwen35-9b-v4 vlmauditor/launch/train_v4.sh

Why no thoughts first: the round-4 thoughts were written from the injected bug's description (privileged), the no-thought file also has the 2,205 round-5 routes, and inference is faster. To try thoughts later, set TRAIN_TRAJ=/data/v4/traj_v4_clean/train_cot_clean.jsonl VAL_TRAJ=/data/v4/traj_v4_clean/val_cot_clean.jsonl; no extra download is needed.

Pick checkpoints by val_traj / val_ident (our held-out worlds), never by the benchmark's 87 tasks.

Downloads last month: 21

Total file size:

111 GB