Datasets:

Languages:
English
ArXiv:
License:
Dataset Preview
Duplicate
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:    CastError
Message:      Couldn't cast
id: string
source: string
domain: string
prompt_ids: list<item: int64>
  child 0, item: int64
completion_ids: list<item: int64>
  child 0, item: int64
finish_reason: string
to
{'id': Value('string'), 'source': Value('string'), 'prompt_ids': List(Value('int64')), 'completion_ids': List(Value('int64')), 'finish_reason': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1835, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 621, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, 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 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              id: string
              source: string
              domain: string
              prompt_ids: list<item: int64>
                child 0, item: int64
              completion_ids: list<item: int64>
                child 0, item: int64
              finish_reason: string
              to
              {'id': Value('string'), 'source': Value('string'), 'prompt_ids': List(Value('int64')), 'completion_ids': List(Value('int64')), 'finish_reason': Value('string')}
              because column names don't match
              
              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 1379, 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 956, 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 1702, 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 1887, 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.

id
string
source
string
prompt_ids
list
completion_ids
list
finish_reason
string
code-gen-00731#0
q36-anchor
[ 248045, 846, 198, 8372, 198, 2523, 513, 17610, 393, 45, 3, 8427, 7552, 11, 28853, 494, 393, 15, 3, 310, 393, 45, 12, 16, 12576, 1946, 15936, 264, 10031, 28750, 11, 488, 1902, 1542, 264, 55968, 13618, 383, 1754, 2357, 6681, 2957, 13,...
[ 760, 3377, 16561, 364, 279, 7776, 2695, 310, 3814, 1396, 2357, 494, 393, 15, 3, 310, 393, 45, 12, 16, 3, 6681, 2957, 11, 5732, 506, 2357, 393, 15, 12576, 561, 5732, 2357, 393, 15, 3, 369, 2582, 328, 32895, 487, 694, 567, 1144, 3...
length
math-ask-00478#0
q36-anchor
[ 248045, 846, 198, 623, 449, 3128, 41520, 20837, 393, 32, 417, 351, 53031, 449, 3128, 41520, 20837, 393, 35, 458, 426, 3, 369, 1583, 16917, 11, 1436, 279, 6851, 13, 561, 16720, 393, 35, 11, 458, 53031, 321, 393, 37, 3, 9940, 383, 2...
[ 8160, 579, 264, 7047, 1817, 310, 11290, 279, 3377, 25, 271, 16, 13, 220, 2972, 15771, 2598, 279, 38622, 64700, 561, 16050, 369, 310, 1423, 279, 11013, 314, 279, 2982, 314, 20837, 393, 39, 22150, 3, 310, 279, 2982, 314, 20837, 393, 6...
length
science-gen-00981#0
q36-anchor
[ 248045, 846, 198, 15542, 378, 3001, 640, 13806, 44312, 278, 70997, 362, 11, 279, 38782, 44751, 507, 7696, 351, 87506, 7697, 318, 1123, 35, 8, 35456, 287, 264, 18392, 722, 11128, 90744, 2227, 40628, 421, 13534, 20716, 14340, 66984, 321, ...
[ 760, 1156, 6587, 310, 10229, 279, 4252, 3874, 314, 279, 29122, 15448, 314, 279, 38782, 44751, 507, 7696, 351, 87506, 7697, 318, 1123, 35, 8, 90744, 2227, 40628, 2261, 378, 3001, 640, 13806, 44312, 278, 70997, 362, 13, 271, 9764, 579, ...
stop
chat-ask-01838#0
q36-anchor
[ 248045, 846, 198, 7734, 264, 11190, 321, 11346, 48271, 10125, 1881, 1330, 9605, 303, 264, 39564, 11, 2583, 20623, 9212, 11, 51067, 3992, 11, 321, 14619, 310, 279, 9605, 6, 7465, 321, 37796, 13, 20495, 279, 9140, 321, 42358, 314, 279, ...
[ 8160, 579, 264, 7047, 7047, 8240, 271, 16, 13, 220, 2972, 1847, 7355, 279, 9640, 64700, 198, 262, 348, 256, 38622, 25, 23502, 11, 11346, 48271, 10125, 1881, 1330, 39564, 9605, 13, 198, 262, 348, 256, 32666, 25, 67137, 9212, 11, 51067,...
stop
chat-ask-01775#0
q36-anchor
[ 248045, 846, 198, 12525, 488, 4052, 1010, 6037, 28007, 54534, 310, 3814, 303, 443, 17776, 12673, 11, 6824, 30, 248046, 198, 248045, 74455, 198, 248068, 198 ]
[ 8160, 579, 264, 7047, 1817, 25, 271, 16, 13, 220, 2972, 15771, 2598, 2570, 5952, 64700, 198, 256, 471, 2972, 25562, 64700, 48023, 54534, 310, 3814, 198, 256, 471, 2972, 4555, 64700, 443, 17776, 12673, 11, 6824, 198, 256, 471, 2972, 15...
stop
if-ask-01633#0
q36-anchor
[248045,846,198,6503,488,3300,449,2560,383,279,13789,17602,310,279,2136,17927,1785,30,271,7525,1965,(...TRUNCATED)
[8160,579,264,7047,1817,25,271,16,13,220,2972,2014,53983,2570,5396,64700,198,256,471,2972,14162,6470(...TRUNCATED)
stop
if-ask-01397#0
q36-anchor
[248045,846,198,4274,279,20438,3559,321,4087,279,3296,466,1732,279,7359,25,10166,557,279,1844,61520,(...TRUNCATED)
[8160,579,264,7047,1817,25,271,16,13,220,2972,2014,53983,2570,5396,64700,198,256,471,2972,14162,6470(...TRUNCATED)
stop
if-ask-01279#0
q36-anchor
[248045,846,198,58387,2923,13909,28253,10895,383,1204,310,1791,264,33404,13205,706,364,264,2902,4314(...TRUNCATED)
[8160,579,264,7047,7047,8240,271,16,13,220,2972,1847,7355,279,1156,579,1622,64700,198,262,348,256,38(...TRUNCATED)
stop
science-gen-01008#0
q36-anchor
[248045,846,198,623,3000,92768,11,279,17312,4649,6463,318,72549,8,1881,3492,314,449,369,4632,258,192(...TRUNCATED)
[760,1156,369,9859,264,3050,3296,883,10922,20665,11,11352,8559,279,32735,12001,53230,318,72549,8,303(...TRUNCATED)
stop
math-ask-00251#0
q36-anchor
[248045,846,198,16,13,264,8,1690,864,2663,314,393,76,3,321,393,77,3,1503,279,7258,271,13682,198,59,3(...TRUNCATED)
[8160,579,264,7047,1817,310,11290,279,3377,25,271,16,13,220,2972,2014,53983,3534,318,64,188912,198,2(...TRUNCATED)
stop
End of preview.

QDA recipe inputs

Everything needed to rebuild the released QDA checkpoints: for each checkpoint, the recipe it was built with and the calibration samples and distillation (KD) traces that recipe reads. A rebuild matches the published checkpoint's KL to BF16 within 5 %; weights are not bit-identical.

Licence: CC BY-NC-SA 4.0, for non-commercial research only, because several folders contain rows from our calibration pools (see "Licences"). Every row also remains subject to the terms of the model or dataset it came from.

Use

hf download qda-lab/qda-recipe-inputs --repo-type dataset --include "QDA-Qwen3.5-4B-W3A16/*" --local-dir inputs
cd inputs/QDA-Qwen3.5-4B-W3A16
qda build qwen3.5-4b-w3a16.yaml --set out=QDA-Qwen3.5-4B-W3A16-rebuild

qda is the QDA library's command-line tool. The recipe's header gives the exact command (some recipes need several GPUs; two folders hold a recipe in several parts, run in order). Run it inside the folder: input paths in a recipe are relative to the recipe file, while out (where the rebuilt checkpoint is written) is relative to the directory you run from. A single-part recipe does not set out, so its command gives it with --set out=...; in a multi-part folder each part's out is set in the recipe, and the next part reads it from the folder.

Layout

One folder per released checkpoint, named after its model repository (qda-lab/<folder>). Each folder holds the recipe (*.yaml, the same build configuration as the released build; only input paths differ, except where a recipe's header says it was changed to fit 80 GB cards), exactly the files it reads, byte-identical to the build, and SHA256SUMS. Files ending in .tokenizer name the tokenizer (sha256 of its tokenizer.json) of the pretokenized file beside them.

folder contents licence
QDA-Qwen3.5-4B-W3A16 GPTQ calibration: 2,802 reasoning traces by Qwen3.5-4B Apache-2.0
QDA-Qwen3.5-4B-W4A16 GPTQ calibration: 3,034 answers by Qwen3.5-4B + 2,048 chat rows from our pools Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3.5-9B-W4A16 GPTQ calibration (1,024 pretokenized rows: traces + text) and 3 rounds of KD traces by Qwen3.5-9B Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3.5-9B-W4A8 GPTQ calibration (512 pretokenized rows: 384 traces + 128 text) and KD traces by Qwen3.5-9B Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3.5-9B-W3A16 the 3-bit start checkpoint (start/, 6.7 GB; our GSQ run, official GSQ code), KD traces for two rounds by Qwen3.5-9B and its 3-bit checkpoints, and the default calibration text Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3.8-27B-W4A16 GPTQ calibration (1,234 pretokenized rows) and KD traces by Qwen3.8-27B Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3.8-27B-W4A8 GPTQ calibration: 600 pretokenized rows, 200 traces by Qwen3.8-27B + 400 default calibration rows Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3.8-27B-W4A16-dsh GPTQ calibration: 768 answers by Qwen3.8-27B to prompts from our pools' reasoning config (own-answers-768.jsonl) + 2,011 rows of our pools' d1 config (d1-filtered.jsonl; the 37 rows removed are listed in d1-filter.json) Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3.8-27B-W3A16 the 3-bit start checkpoint (start/, 14.6 GB): ISTA-DASLab/Qwen3.8-27B-3Bit-GSQ with a BF16 embedding and the MTP block added (derive_start.py rebuilds it from ISTA's release; see start/README.md), KD traces (2,167, by Qwen3.8-27B) and the default calibration text Apache-2.0 (start: ISTA-DASLab); CC BY-NC-SA 4.0
QDA-Qwen3.6-35B-A3B-W4A8 GPTQ calibration: 4,096-token pretokenized mix of Qwen3.6-35B-A3B's own answers and default text Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3.6-35B-A3B-W4A16 GPTQ calibration (7,326 pretokenized 4,096-token rows: own answers + default text) and three KD trace sets by Qwen3.6-35B-A3B (6,082 + 1,600 + 3,619) Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3.6-35B-A3B-W3A16 the 3-bit start checkpoint (start/, 17.3 GB; our GSQ run, official GSQ code), four KD trace sets by Qwen3.6-35B-A3B and its 3-bit checkpoints, the default calibration text, and four recipes run in order: -ace, then -h and -i, then -soup Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3-Coder-30B-A3B-W4A16 6,050 traces by Qwen3-Coder-30B-A3B (token ids and chat form) + chat and text rows from our pools Apache-2.0; CC BY-NC-SA 4.0
QDA-Gemma-4-26B-A4B-W4A16 GPTQ calibration: 6,308 answers by Gemma 4 26B-A4B Apache-2.0
QDA-Gemma-4-31B-W4A16 GPTQ calibration text, 2,120 KD traces by Gemma 4 31B, and the contamination index (prompt hashes; source paths reduced to file names) the calibration is screened with Apache-2.0; CC BY-NC-SA 4.0
QDA-Nemotron-3-Nano-30B-A3B-W4A16 3,498 answers by Nemotron 3 Nano + chat and text rows from our pools NVIDIA Open Model License; CC BY-NC-SA 4.0
QDA-Nemotron-3-Nano-30B-A3B-W3A16 a 12-part GPTQ calibration mix: answers by Nemotron 3 Nano + chat, text and multilingual rows from our pools NVIDIA Open Model License; CC BY-NC-SA 4.0
QDA-DeepSeek-R1-Distill-Llama-8B-W4A16 GPTQ calibration (default text, packed to 2,048 tokens) and two rounds of KD traces by the model MIT (Llama 3.1 base); CC BY-NC-SA 4.0
QDA-DeepSeek-R1-Distill-Llama-8B-W4A8 GPTQ calibration (114 packed 2,048-token blocks) and the KD traces of its three distillation parts (6,100 + 2,739 + 939), and four recipes run in order: part 1, then -long2 and -long, then -soup MIT (Llama 3.1 base); CC BY-NC-SA 4.0
QDA-DeepSeek-R1-Distill-Llama-8B-W3A16 GPTQ calibration: our pools' d1 config in this model's chat template (34 rows removed) + 1,913 answers by the base model to prompts from our pools' reasoning, chat and multilingual configs (3 of them code_contests prompts not in our pools) (calib_mix.jsonl); KD: those answers + 1,703 answers by an intermediate quantized checkpoint of the same search (kd_mix.jsonl, token ids); the prompts are in prompts_all.jsonl MIT (Llama 3.1 base); CC BY-NC-SA 4.0; Apache-2.0; CC BY 4.0 (code_contests); CC BY-SA 4.0
QDA-Qwen3.8-27B-Q6-GGUF the GGUF recipe (needs a llama.cpp checkout, see its header) and its calibration: 1,234 prompts with Qwen3.8-27B's own answers as token ids (kd-traces-1234.jsonl) and the prompts as text (kd-prompts-1234.jsonl); the build used 1,226 of them, after its contamination filter removed 8 (row numbers in kd-filter.json) Apache-2.0; CC BY-NC-SA 4.0; CC BY 4.0 (9 code_contests prompts)
QDA-Qwen3.6-35B-A3B-Q4-GGUF the GGUF recipe (needs a llama.cpp checkout, see its header) and its calibration: the 2,048 rows of our pools' chat config (chat-2048.jsonl) and the build's contamination filter record (chat-filter.json) CC BY-NC-SA 4.0

A start/ subfolder holds the 3-bit checkpoint a recipe starts from (model: ./start); run those recipes from inside their folder. QDA-Qwen3.6-35B-A3B-W4A16-SGLang is a repack of QDA-Qwen3.6-35B-A3B-W4A16 and uses its folder.

Licences

  • Traces and answers written by a model follow that model's licence: Qwen3.5, Qwen3.6, Qwen3.8 and Qwen3-Coder: Apache-2.0; Gemma 4: Apache-2.0; NVIDIA Nemotron 3 Nano: NVIDIA Open Model License; DeepSeek-R1-Distill-Llama-8B: MIT, on a Llama 3.1 base (Llama 3.1 Community License).
  • Chat, text and multilingual rows from our calibration pools, and the default calibration text, inherit the licence of the QDA calibration pools: CC BY-NC-SA 4.0 (some of their sources are non-commercial).
  • The 3-bit start checkpoints in start/ follow Apache-2.0. QDA-Qwen3.8-27B-W3A16/start/ is ISTA-DASLab/Qwen3.8-27B-3Bit-GSQ by IST Austria's DAS Lab (Apache-2.0; the licence text and our changes are in that folder); the credit for that quantization is theirs. The 9B and 35B starts are our own runs of the official GSQ code (IST-DASLab/GSQ, paper) on the Apache-2.0 Qwen models.
  • Prompts come from public datasets; each row stays subject to its source's terms.

Notes

  • Self-generated data is included as files only; generating it again is not part of this release.
  • Some calibration text was screened against evaluation prompts and some was not: do not use these files to calibrate a model you will evaluate on the same benchmarks without checking overlap first.
  • Check a folder with sha256sum -c SHA256SUMS.

NVFP4 W4A4 recipes

The five NVFP4 W4A4 model releases each have a recipe folder here:

folder contents licence
QDA-DeepSeek-R1-Distill-Llama-8B-W4A4 the two-stage W4A4 recipe (GPTQ on 256 × 4,096-token samples, then 100 scale_kd steps), its trace input (1,948 traces as token ids, 1,092 by the model and 856 by an earlier checkpoint of the run), the generator that made it (tools/gen_traces.py) and the three pool files its runs read MIT (Llama 3.1 base); CC BY-NC-SA 4.0
QDA-Qwen3.5-9B-W4A4 the W4A4 recipe (GPTQ on 494 × 4,096 tokens) and its calibration: the W4A8 mix's rows with 18 MMLU-Pro hits dropped, and the guard script that dropped them Apache-2.0; CC BY-NC-SA 4.0
QDA-Gemma-4-26B-A4B-W4A4 the W4A4 recipe (GPTQ on 128 × 2,048 tokens, router-weighted expert Hessians) and its calibration mix (6,308 rows of text, the same file the W4A16 release of this model was built on) Apache-2.0
QDA-Qwen3.6-35B-A3B-W4A4 the W4A4 recipe (head, then GPTQ on 384 × 2,048 tokens) and its calibration: the self-mix-4k mix with 36 MMLU-Pro rows dropped, and the guard script that dropped them Apache-2.0; CC BY-NC-SA 4.0
QDA-Qwen3.8-27B-W4A4 the W4A4 recipe (GPTQ on 52 × 4,096 tokens) and its calibration: 162 answers by the model to our own 162 prompts, plus those prompts and the script that wrote them Apache-2.0 (answers); CC BY-NC-SA 4.0 (our prompts)

Each of the five W4A4 folders also holds CHECKPOINT.sha256, the sha256 of every file of the released checkpoint, weights included; the weights themselves are in the model repository qda-lab/<folder>, not here. SHA256SUMS covers the files in the folder.

The five W4A4 recipes set out to the release name themselves, so no --set out=... is needed:

hf download qda-lab/qda-recipe-inputs --repo-type dataset --include "QDA-Qwen3.8-27B-W4A4/*" --local-dir inputs
cd inputs/QDA-Qwen3.8-27B-W4A4
qda build qwen3.8-27b-w4a4-nvfp4.yaml

A rebuild's KL to BF16 is within ±5 % of the published checkpoint's; the weights are not bit-identical.

Citation

QDA: Closed-Loop LLM Quantization Design with Agents

@article{li2026qda,
  title   = {QDA: Closed-Loop LLM Quantization Design with Agents},
  author  = {Li, Xingyang and Zhao, Tianlang and Han, Song and Zhu, Ligeng},
  journal = {arXiv preprint arXiv:2611.23388},
  year    = {2026}
}
Downloads last month
492

Papers for qda-lab/qda-recipe-inputs