How Does Calibration Data Affect the Post-training Pruning and Quantization of Large Language Models?
Paper • 2311.09755 • Published
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 290, in _generate_tables
pa_table = paj.read_json(
io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
)
File "pyarrow/_json.pyx", line 342, 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 string 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 304, 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.
MIX-STQ quantization research artifacts
Importance matrices and measured accuracy results from a study of low-bit quantization and the evidence boundary between a proxy encoder and a storable format. Source and full research records: https://github.com/Topabaem05/mix-stq
Why these files exist
An importance matrix takes GPU time to collect and is easy to lose. The Qwen3.8-27B matrix here was recollected twice after being dropped, at about five minutes of 96 GB GPU time each. It is pinned to an exact model revision, so it is reusable rather than something to regenerate.
Contents
| File | What it is |
|---|---|
qwen38_imatrix.json |
Per-channel E[x^2] for all 192 MLP tensors of Qwen3.8-27B, plus metadata |
qwen38_imatrix.pt |
Same data as torch tensors, keyed by module name |
qwen38_frontier.json |
Historical MMLU+ARC sweep using approximate, non-storable IQ encoders |
reference-v25/qwen38_reference.json |
Reference-constrained IQ3 result with all paired correctness vectors |
reference-v25/qwen38_reference.log |
Complete 192-tensor quantization and evaluation log |
reference-v25/mix-stq-v25-reference-iq3.md |
Corrected interpretation and next research gates |
reference-v25/qwen38_top1.svg |
Flint chart separating proxy and reference arms |
olmoe_imatrix.json |
Importance matrix for OLMoE-1B-7B-0924 expert tensors |
frontier4.json, frontier5.json |
Eight-point bit-budget frontier on OLMoE |
task_accuracy.json |
Paired task-accuracy comparison, learned vs fixed codebook |
Collection settings
| Setting | Value |
|---|---|
| Model | Qwen/Qwen3.8-27B @ 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 |
| Samples | 96 documents, 2048 tokens each |
| Domains | wiki, code, chat (stratified, three sources) |
| Observed tokens | 23,553 |
| Tensors | 192 (64 layers x gate_proj, up_proj, down_proj) |
| Statistic | Per-input-channel mean of x squared, via forward hooks |
Stratifying across three domains rather than one avoids the calibration dependence reported in arXiv:2311.09755.
Headline measurement
Qwen3.8-27B, MMLU 140 + ARC-Challenge 60, paired McNemar plus bootstrap CI. The measured dense baseline was FP16, not the BF16 model-card result.
| Arm | Encoder | bpw | Accuracy | vs dense FP16 | p |
|---|---|---|---|---|---|
| dense FP16 | dense | 16 | 0.8050 | baseline | — |
| IQ2_XXS | approximation, not storable | 2.0625 | 0.7300 | −0.0750 | 0.0007 |
| IQ3_XXS | approximation, not storable | 3.0625 | 0.7900 | −0.0150 | 0.5078 |
| IQ3_XXS | reference-constrained | 3.0625 | 0.7700 | −0.0350 | 0.1185 |
| IQ3_S | approximation, not storable | 3.4375 | 0.7900 | −0.0150 | 0.4531 |
The reference IQ3 point estimate is 3.5 percentage points below dense FP16. Its paired 95% interval permits a dense advantage from 0.0 to 7.5 points, so the result establishes neither significant damage nor equivalence. The former full-accuracy claim is withdrawn pending an 800+ item, dtype-aligned run.
What these numbers are not
Accuracy was measured by quantizing weights in PyTorch and scoring the model, not by writing a GGUF file and running llama.cpp. GGUF round-trip and C decoder parity are unverified for the IQ tiers. There is no quantized model checkpoint in this repository, only the importance data and the measurements. IQ2 still has no reference encoder, and only the 192 MLP tensors were quantized.
- Downloads last month
- 125