Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 82, in _split_generators
                  first_examples = list(islice(pipeline, self.NUM_EXAMPLES_FOR_FEATURES_INFERENCE))
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 58, in _get_pipeline_from_tar
                  current_example[field_name] = cls.DECODERS[data_extension](current_example[field_name])
                                                ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 336, in torch_loads
                  return torch.load(io.BytesIO(data), weights_only=True)
                         ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/torch/serialization.py", line 1609, in load
                  raise pickle.UnpicklingError(_get_wo_message(str(e))) from None
              _pickle.UnpicklingError: Weights only load failed. This file can still be loaded, to do so you have two options, do those steps only if you trust the source of the checkpoint. 
              	(1) In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
              	(2) Alternatively, to load with `weights_only=True` please check the recommended steps in the following error message.
              	WeightsUnpickler error: Unsupported global: GLOBAL numpy.core.multiarray._reconstruct was not an allowed global by default. Please use `torch.serialization.add_safe_globals([numpy.core.multiarray._reconstruct])` or the `torch.serialization.safe_globals([numpy.core.multiarray._reconstruct])` context manager to allowlist this global if you trust this class/function.
              
              Check the documentation of torch.load to learn more about types accepted by default with weights_only https://pytorch.org/docs/stable/generated/torch.load.html.
              
              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/split_names.py", line 68, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

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.

LightText Data and Pure Four-Layer Runtime Assets

Private owner-controlled backup of the data inputs, pretrained assets and one completed pure-four-text-layer baseline checkpoint. Keep this repository private. Environment packages and compiled extensions are not included; portability on a new device has not been execution-tested.

Included Data

  • Training: 562 scenes and 36,665 referring expressions.
  • Validation/evaluation: 141 scenes and 9,508 expressions. The published test entry uses this validation split, not a separate official test split.
  • 703 preprocessed point-cloud tuples at scannetv2/{train,val}/*_refer.pth, containing XYZ, RGB, superpoints, semantic labels and instance labels.
  • 703 required CLIP features at clip_feat/<scene_id>.pth.
  • ScanRefer language annotations, precomputed language graphs, lookup.json, and scannetv2-labels.combined.tsv.

These point clouds are the existing training-ready preprocessing, not original .sens/RGB-D sensor recordings. Original sensor scans were not present in the selected training source.

Pretrained and Test Assets

  • runtime-pretrained.tar.gz: the exact existing roberta-base/ directory files and sp_unet_backbone.pth. The legacy fast-tokenizer JSON embeds its BPE merge rules; the source did not contain a separate merges.txt, and no replacement model was downloaded.
  • runtime-pure4-checkpoint.tar.gz: completed pure-four-text-layer baseline runtime_assets/pure4/best.pth, its untouched original YAML, a path-only configuration copy at runtime_assets/configs/model_base_early_exit_4.yaml, and runtime_assets/ASSET_NOTICES.md.
  • This checkpoint is for framework.model.model_base_early_exit_4, not for every historical experimental variant. Its seven model source files match the published code snapshot byte-for-byte.
  • Original source hashes match the previously recorded completed-baseline export binding. No checkpoint was deserialized and no new model evaluation or training was run for this backup.

Restore

Code: https://github.com/songchuanle-1/lighttext-project . The matched published snapshot commit is 9681568298555d165c7872059a638140ecde935c.

Download all 43 archives in UPLOAD_MANIFEST.json. Verify compressed-file SHA-256 values, then extract every archive into the same <cloned-code-repository>/data/ directory. Existing data archives keep their original scannetv2, scanrefer, and clip_feat paths. The two runtime archives add:

data/
  roberta-base/
  sp_unet_backbone.pth
  runtime_assets/
    configs/model_base_early_exit_4.yaml
    pure4/best.pth
    pure4/original_model_base_early_exit_4.yaml
    ASSET_NOTICES.md

The path-only YAML expects commands to run from the cloned repository root, with data in ./data and output in ./checkpoints. Model, seed, schedule, loss, and dataset fields are otherwise unchanged from the checkpoint's original YAML. Keep the configuration basename unchanged: the native loader selects the model from its name (or FTE_MODEL).

After independently preparing a compatible environment, the native Linux-shell test entry is:

FTE_MODEL=model_base_early_exit_4 python -B tools/test.py data/runtime_assets/configs/model_base_early_exit_4.yaml --checkpoint data/runtime_assets/pure4/best.pth --num_gpus 1 --gpu_ids 0

The corresponding training entry, initialized from the included pretrained backbone, is:

FTE_MODEL=model_base_early_exit_4 python -B tools/train.py data/runtime_assets/configs/model_base_early_exit_4.yaml --num_gpus 1 --gpu_ids 0

These are native entry commands, not a claim of new-device execution or exact cross-environment score reproduction. Native profiling defaults are unchanged. No environment installation, compiled binaries, running-job checkpoints, private logs, credentials, or checkpoints for other experimental variants are bundled.

Each archive includes a per-file MANIFEST.json. CLIP_MANIFEST.json records feature provenance; RUNTIME_MANIFEST.json records the pretrained/checkpoint checks and exact configuration changes. Python/PyTorch .pth and .bin files can contain executable pickle data: load only trusted, hash-verified files.

Provenance and Terms

The original data and weights were streamed read-only; training processes, original configurations, source files and schedules were not changed.

ScanNet: Angela Dai et al., CVPR 2017, https://kaldir.vc.in.tum.de/scannet/ScanNet_TOS.pdf . ScanRefer: Dave Zhenyu Chen et al., ECCV 2020; its language annotations have separate CC BY-NC-SA 3.0 requirements, which do not replace ScanNet's geometry terms.

RoBERTa model card: https://huggingface.co/FacebookAI/roberta-base (declares MIT). Copyright (c) Facebook, Inc. and its affiliates; upstream license: https://github.com/facebookresearch/fairseq/blob/main/LICENSE . IPDN identifies the point backbone as SPFormer: https://github.com/80chen86/IPDN#pretrained-backbone . See the archived asset notices for component-specific provenance.

No new blanket license is granted for the datasets, derived features, pretrained assets or inherited model components. Do not make this backup public without independently addressing redistribution rights.

Downloads last month
31