The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
task: string
metric: string
floor_pp: double
sigma: double
n: int64
eval: string
arm_desc: string
within_run: bool
decoding: string
serve: string
engine: string
arm: string
value: double
experiment: string
to
{'experiment': Value('string'), 'task': Value('string'), 'metric': Value('string'), 'arm': Value('string'), 'value': Value('float64'), 'n': Value('int64'), 'within_run': Value('bool'), 'engine': Value('string'), 'eval': Value('string'), 'decoding': Value('string'), 'serve': Value('string'), 'arm_desc': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
task: string
metric: string
floor_pp: double
sigma: double
n: int64
eval: string
arm_desc: string
within_run: bool
decoding: string
serve: string
engine: string
arm: string
value: double
experiment: string
to
{'experiment': Value('string'), 'task': Value('string'), 'metric': Value('string'), 'arm': Value('string'), 'value': Value('float64'), 'n': Value('int64'), 'within_run': Value('bool'), 'engine': Value('string'), 'eval': Value('string'), 'decoding': Value('string'), 'serve': Value('string'), 'arm_desc': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Quantization Rebuild Noise Floor
Running the same quantization recipe twice produces two checkpoints that differ by more than most papers' reported deltas. This dataset is the measurement.
We quantized Qwen3.8-27B to W4A16 with GPTQ, then ran the exact same recipe a second time —
same model, same settings, same calibration set, only a different quantization run. We evaluated both
builds in a single serving run so no engine or configuration difference could leak in, then measured
how far apart they landed.
| Task | Metric | Gap between two identical-recipe builds |
|---|---|---|
| GSM8K (en) | exact_match, strict | 3.56 pp (1.83σ, n=1,319) |
| GSM8K (en) | exact_match, flexible | 2.27 pp (1.18σ) |
| GSM8K (ko) | exact_match, strict | 2.13 pp (1.10σ) |
| GSM8K (ko) | exact_match, flexible | 1.37 pp (0.73σ) |
| RAG, absent-information | fabrication rate | 27.08 pp (7.5σ, n=240) |
Why this matters
Quantization work routinely compares checkpoints built once each — recipe A versus recipe B, one build apiece — and attributes the difference to the recipe. If the build-to-build spread is larger than the observed gap, that attribution does not hold.
Our own work ran into this. We had concluded that one of our checkpoints hallucinated far more than a third-party one on absent-information questions (9.17% vs 1.25%). Then we measured the floor: a second build of the same recipe scored 36.25%. The gap we were about to ship a decision on was smaller than the noise.
A practical rule that falls out of this: if the observed difference is under 2× the rebuild floor for that metric, do not attribute it to the treatment. Report it as unresolved and build more copies.
The second finding: accuracy metrics hide it
The bad build was not bad everywhere. ours-r2 had the highest extraction accuracy of all four arms
(0.5375) and an identical grounded-answer score. Only abstention behaviour — refusing to answer when the
document does not contain the value — collapsed. A quality gate built on accuracy alone would have
shipped it.
Contents
data/measurements.jsonl— per-arm, per-metric results with the serving configuration attached. Arms:ours(build #1),ours-r2(identical recipe, second run),ours-mamba(third build with a different ignore list),bf16(unquantized),int4-redhat(third-party frozen artifact).data/rebuild_noise_floor.jsonl— the derived floor per task and metric.
Method
Both GSM8K arms ran in one serving session (vLLM 0.27.1, lm-eval 0.4.12, greedy decoding, full
1,319-item sets, max_model_len=8192, max_num_seqs=64). The RAG arms ran in a separate single session
(max_model_len=131072, max_num_seqs=256, n=240 synthetic Korean procurement documents at hard
difficulty). Serving configuration was identical across arms within each session; only the model path
changed. Do not compare absolute values across the two sessions — the configurations differ.
Limits
- One pair. Two builds give one difference. That is a lower bound on the spread, not a standard deviation. Estimating the distribution needs more copies.
- Floors are per metric and per task. The 3.56 pp figure does not transfer to other benchmarks.
- This is our pipeline. Frozen third-party checkpoints are shipped artifacts; a customer receives that exact file, so build variance is not part of their decision. It is part of ours, because we rebuild.
ours-mambais a third build with a different ignore list, so its gap mixes treatment and build noise. It is included for completeness, not as a controlled comparison.- One model family, one quantization method, one hardware generation.
Citation
@misc{thakicloud2026rebuildfloor,
title = {Quantization Rebuild Noise Floor},
author = {ThakiCloud},
year = {2026},
url = {https://huggingface.co/datasets/ThakiCloud/quantization-rebuild-noise-floor}
}
- Downloads last month
- -