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/packaged_modules/json/json.py", line 291, in _generate_tables
pa_table = paj.read_json(
io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
)
File "pyarrow/_json.pyx", line 364, in pyarrow._json.read_json
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: JSON parse error: Column() changed from object to number in row 0
During handling of the above exception, another exception occurred:
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/json/json.py", line 101, in _split_generators
pa_table = next(iter(self._generate_tables(**splits[0].gen_kwargs, allow_full_read=False)))[1]
~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 306, in _generate_tables
batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
examples = [ujson_loads(line) for line in original_batch.splitlines()]
~~~~~~~~~~~^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
return pd.io.json.ujson_loads(*args, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
ValueError: Expected object or value
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.
Flare SM70 FP32 accuracy references
FP32 reference logits for the accuracy (full-vocabulary KL / top-1) export gate of the
SM70 (Volta, V100) port of Flare:
github.com/Atlas3DSS/flare-local-inference
(branch atlas/sm70-public).
Making these files yourself takes an FP32 forward pass of the 27B checkpoint, and the long
reference needs a 96 GB GPU. With these files, someone who has only one V100 can still run every
export gate. See docs/sm70/EXPORT.md, section "One V100 only", and tools/fetch_references.sh
in the repository, which downloads this dataset and checks every SHA-256 below.
Files
| File | Bytes | SHA-256 |
|---|---|---|
reference.safetensors |
140,054,888 | 30dd669b1baf6a1ec517b635c083b275dd72417ca07521bf33cecdd954f02a3f |
reference.json |
6,590 | 277f47825bac44feba200dfab6ab555d637b66a30d31c2bfdfa2f0f09afcc5a3 |
long-code-reference.safetensors |
286,266,648 | d4479957d5ec101c8274d7a7fe59032d0171423ebacd8607a6aeb61e00f1afc9 |
long-code-reference.json |
343,709 | 663ab9262b33ae26a4f07dbbd08ac02ea188990dbc1cc7c8df1effa474e0f15a |
prompts.jsonl |
87,395 | b200be2010c1c4eb24b7cd787141e7db23cefc7df7cc41e2a232764a92aad134 |
SHA256SUMS lists the same digests in sha256sum -c format.
reference.safetensors: three 32-token prompts with 16 greedy tokens each. It holds the prompt and generated ids, plus FP32 prefill logits[32, 248320]and decode logits[15, 248320]for each prompt.long-code-reference.safetensors: the six longest non-thinking coding prompts ofprompts.jsonl(about 4,100 tokens each, rendered with the checkpoint's chat template,enable_thinking=False), each followed by a 32-token greedy continuation. It stores FP32 logits[48, 248320]at the last 48 positions of one teacher-forced forward pass without a cache, with the matching input ids and positions.- The two
.jsonfiles are the provenance reports written next to each artifact. The accuracy gate reads them: it checks the model ID, the revision, the checkpointconfig.jsonandmodel.safetensors.index.jsonhashes, and the artifact digest. Keep each one next to its.safetensorsfile. They are published byte for byte as written, so the digests recorded in the repository's receipts (docs/sm70/receipts/measurements/reference-kl.json) match. They name the generating script under its earlier file name, and the paths in them are the paths on the generating host. prompts.jsonl: the public generated input setpublic-generated-v1-new-baselinethat the long reference was cut from.tools/generate_prompts.py --checkpoint <Qwen3.8-27B-NVFP4> --target-tokens 4096reproduces it byte for byte. It is included only for cross-checking.
How they were made
- Model:
nvidia/Qwen3.8-27B-NVFP4at revision482ca0f3832238542f8f5295dde86b5f22711d80. Checkpointconfig.jsonsha25671fa01bb64de20971045cd82a82c0ebb25a54e1b0a3f09ef23cf5479e9c54df8,model.safetensors.index.jsonsha2567aa103a2582b7d26631988de33dea19e8a308ee9c239e8e14feb374af30905e2. - Arithmetic: the canonical Hugging Face Transformers forward in FP32 (eager attention), with the raw NVFP4 weights decoded weight-only to FP32 and the checkpoint's calibrated FP8 activations applied. No Flare kernels were involved.
- Environment: torch
2.14.1+cu130, transformers5.21.0.dev0, one NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB). - Scripts, from the repository:
tools/reference_model.py(sha256c54f7c2dcaf113ed651b0fcea684a99e7c5a7dba847e9ce8a0afc027719bb9e5, unchanged since) and the GPU-resident wrapper now namedtools/gpu_reference.py(sha256 at generationc116406c73c55fab08d6094cc1e95933c3c00dcfab57466e6f738e9ae5ff5ed4. It was renamed later, and only its docstring, messages and one provenance key name changed). The wrapper copies the raw checkpoint to the GPU and runsreference_model.py's own decode and forward there. Its--self-checkconfirms that the decoded weights are bit-identical to the CPU decode. - Commands (the paths are placeholders):
python tools/generate_prompts.py --output refs/prompts.jsonl \
--checkpoint Qwen3.8-27B-NVFP4 --target-tokens 4096
python tools/gpu_reference.py repro --root refs --out refs --calib-out refs \
--checkpoint Qwen3.8-27B-NVFP4 --prompts refs/prompts.jsonl --threads 30 --vram-fraction .39
python tools/gpu_reference.py long --root refs --out refs --calib-out refs \
--checkpoint Qwen3.8-27B-NVFP4 --prompts refs/prompts.jsonl --threads 30 --vram-fraction .39 \
--count 6 --new-tokens 32 --positions 48
repro runs reference_model.py's main unchanged (its prompts, batch and greedy loop) and
writes reference.{safetensors,json}. long writes long-code-reference.{safetensors,json}.
Use
tools/fetch_references.sh /data/flare-refs
Then pass --short-reference /data/flare-refs/reference.safetensors and
--long-reference /data/flare-refs/long-code-reference.safetensors to tools/prepare_replay.sh
as docs/sm70/EXPORT.md describes. The gate reports the measured KL and top-1 agreement;
complete means the measurement ran, not that a numerical threshold was passed.
License
Apache-2.0, the license of the model the logits were computed from.
- Downloads last month
- 28