Datasets:

TitanMamo10
/

k2-quant-eval

Error code:   DatasetGenerationCastError
Exception:    DatasetGenerationCastError
Message:      An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 6 new columns ({'test_imports', 'task_id', 'test_list', 'source_file', 'code', 'prompt'}) and 1 missing columns ({'text'}).

This happened while the json dataset builder was generating data using

hf://datasets/TitanMamo10/k2-quant-eval/eval/sanitized-mbpp.json (at revision dc41a10954ed9793e440c95559187b45aa8a4520), ['hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_all_427.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_f16.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_f16_15b.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/sanitized-mbpp.json'], ['hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_all_427.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_f16.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_f16_15b.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/sanitized-mbpp.json']

Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1856, in _prepare_split_single
                  writer.write_table(table)
                  ~~~~~~~~~~~~~~~~~~^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 811, in write_table
                  self._write_table(pa_table, writer_batch_size=writer_batch_size)
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 822, in _write_table
                  pa_table = table_cast(pa_table, self._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
              source_file: string
              task_id: int64
              prompt: string
              code: string
              test_imports: list<item: string>
                child 0, item: string
              test_list: list<item: string>
                child 0, item: string
              to
              {'text': Value('int64')}
              because column names don't match
              
              During handling of the above exception, another exception occurred:
              
              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 1858, in _prepare_split_single
                  raise DatasetGenerationCastError.from_cast_error(
                  ...<4 lines>...
                  )
              datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
              
              All the data files must have the same columns, but at some point there are 6 new columns ({'test_imports', 'task_id', 'test_list', 'source_file', 'code', 'prompt'}) and 1 missing columns ({'text'}).
              
              This happened while the json dataset builder was generating data using
              
              hf://datasets/TitanMamo10/k2-quant-eval/eval/sanitized-mbpp.json (at revision dc41a10954ed9793e440c95559187b45aa8a4520), ['hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_all_427.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_f16.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_f16_15b.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/sanitized-mbpp.json'], ['hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_all_427.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_f16.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/ids_f16_15b.json', 'hf://datasets/TitanMamo10/k2-quant-eval@dc41a10954ed9793e440c95559187b45aa8a4520/eval/sanitized-mbpp.json']
              
              Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)

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.

text int64
2
3
4
6
7
8
9
11
12
14
16
17
18
19
20
56
57
58
59
61
62
63
64
65
66
67
68
69
70
71
72
74
75
77
79
80
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
108
109
111
113
115
116
117
118
119
120
123
124
125
126
127
128
129
130
131
132
133
135
137
138
139
140
141
142
143
145
160
161
162
163
164
165
166
167
168

End of preview.

K2 quantization: evaluation sets, harness and calibration mixes

This dataset repo is the missing data behind the numbers we quote for K2, our EXL3-class trellis quantization (≈3.34 bits/weight) of the Qwen2.5 family. It exists because a maintainer objected — correctly — that our benchmark claims referenced "private datasets not published anywhere". Everything quoted below is now either in this repo, fetched by a script in this repo, or explicitly flagged as not reproducible by a third party (see Caveats, item 6).

It contains token ids and code, not text and not weights.

What is in here

path	what it is
`eval/ids_all_427.json`	all 427 MBPP (sanitized) `task_id`s — the neutral eval set
`eval/ids_f16.json`	180 MBPP ids that Qwen2.5-0.5B F16 solves (the 0.5B subset)
`eval/ids_f16_15b.json`	246 MBPP ids that Qwen2.5-1.5B F16 solves (the 1.5B subset)
`eval/sanitized-mbpp.json`	MBPP sanitized, byte-identical to upstream (Apache-2.0, see `PROVENANCE.md`)
`harness/k2_task_bench.py`	ARC-Easy / GSM8K / HellaSwag / probes battery over any OpenAI-compatible server
`harness/k2_code_bench.py`	MBPP executable bench, pass@2 with one feedback retry
`harness/k2_split_report.py`	the all-427 / F16-solved / F16-failed split report
`calibration/*.npz`	7 token-id mixes used for SC, the calibration trace and scale-QAT (≈175 MiB)
`pipeline/k2_pipeline.py`, `pipeline/k2_pipeline.sh`	the quantization driver: f16 → pack → QAT → apply → GGUF → PPL
`fetch/`	how to fetch the upstream text we do not redistribute (WikiText-2)
`RUN.md`	the exact commands that regenerate a measured number
`PROVENANCE.md`	per file: origin, license, sha256, tokenizer
`LICENSE`, `LICENSE-DATA.md`, `licenses/`	code license, data-license map, Apache-2.0 text

All input_ids arrays are int32, shape (N, 512), Qwen2.5 tokenizer, vocab 151,936 (observed max id 151,664). They are not portable to another tokenizer — see PROVENANCE.md.

What this is not

Not model weights. No GGUF, no safetensors, no tensors. The artifacts themselves live in the model repos (TitanMamo10/Qwen2.5-1.5B-K2-3.34bpw, TitanMamo10/qwen25-qat-q3-poc).
Not a leaderboard. The batteries are small, run on one RTX 3060, and the subsets are F16-selected (i.e. biased toward the reference model). See the caveats.
Not a redistribution of benchmark text. MBPP is the single exception, and only because its Apache-2.0 license explicitly permits it (verified byte-identical to upstream).
Not a general-purpose calibration set. These mixes are tuned for Qwen2.5 at 3.34 bpw on this box. WikiText is deliberately the anchor; that is a deployment choice, and we measured that it costs coding performance (PROVENANCE.md, "what we learned about these mixes").

Numbers these files produce

Measured on one box (RTX 3060 12 GB, 6 threads), llama-perplexity -c 512 --chunks 20 = 10,240 tokens of wiki.test.raw, greedy (temperature 0) for the task battery.

artifact	PPL (wiki.test)	vs its F16	source
Qwen2.5-1.5B F16	10.3099	—	`pipeline_qwen15_clean/base-f16.gguf`
Qwen2.5-1.5B K2 3.34 bpw, clean PTQ	10.5895	+2.71%	`pipeline_qwen15_clean/k2_ptq.gguf` (KLD 0.0485, top-1 89.22%, RMS Δp 6.14%)
Qwen2.5-0.5B K2 3.34 bpw, raw PTQ	15.3825	—	`pipeline_qwen05_*/k2.gguf`
Qwen2.5-0.5B K2 3.34 bpw + Adam scale-QAT	14.7148	−4.34% vs raw PTQ	same pack, new `scales.pt`

Task battery (0.5B on the 180-item F16 subset, 1.5B on the 246-item F16 subset):

task	0.5B F16 → PTQ	1.5B F16 → PTQ
ARC-Easy	68.5 → 65.5	87.0 → 87.0
GSM8K (4-shot)	16.5 → 4.0	15.0 → 38.5
HellaSwag (4-way)	35 → 28	58.0 → 51.0
probes (ours, 100 items)	85 → 70	83.7 → 85.7
MBPP (pass@2 + 1 retry)	100 → 53.3 (n=180)	— → 73.6 (n=246)

The 100/— cells are by construction: the subsets are defined as "items the plain F16 solves", so F16 scores 100% on them and 1.5B has no all-subset F16 number in the log. Always quote all-427 alongside (harness/k2_split_report.py prints all three columns).

Caveats (read before quoting anything)

Subset bias. ids_f16*.json are F16-solved subsets. Raw accuracy on them is biased toward the model that defined them. Report all-427, 246-subset and 181-residual together.
MBPP is pass@2. One feedback retry with the failing assert appended. This is not HumanEval-style pass@1 and must not be compared to pass@1 numbers.
GSM8K is format-sensitive. The extractor takes the first #### marker; a model that never emits the stop token drifts and the fallback rule measures formatting, not arithmetic. The harness records the rule per item (rule field) so this is auditable.
PPL is 20 chunks. 10,240 tokens, -c 512. Long-context behaviour and KLD are separate measurements with their own references.
n is small. GSM8K 100 items on 0.5B, probes 100, HellaSwag 200. Single run, greedy. Differences under ~3 points are not resolved.
One number is not third-party reproducible. Our "technical" and "domain" PPL columns are measured on an internal engineering document, and the "chat"/"alpaca" OOD corpora are held-out slices we assembled ourselves and cannot redistribute. Those columns are flagged in fetch/README.md; treat them as internal-only evidence, not as published benchmarks.
The mixes inherit upstream licenses. WikiText blocks are CC BY-SA 4.0; GSM8K MIT; the chat and ubergarm material has no established license (LICENSE-DATA.md). If you need a clean single license, use poolv3b65.npz minus the chat/ubergarm shares and re-derive — the ratios are documented.
The trainer is not in this repo. pipeline/ holds the stage driver; the scale-QAT trainer (k2_block_qat_v5.py) and the trellis codec depend on our local tree and are shared on request. The measured battery numbers need no trainer: harness/ + eval/
- any OpenAI-compatible server reproduce them.

Citation

@misc{k2_quant_eval_2026,
  author       = {TitanMamo},
  title        = {K2 quantization: evaluation sets, harness and calibration mixes (Qwen2.5, 3.34 bpw)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/TitanMamo10/k2-quant-eval}}
}

License

Code (harness/, pipeline/): MIT — see LICENSE. Data: mixed, inherited from the upstream sources; the per-file map is LICENSE-DATA.md, the derivation is PROVENANCE.md. MBPP text: Apache-2.0 (licenses/Apache-2.0.txt).

Downloads last month: 66

Total file size:

184 MB