The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
benchmark_revisions: struct<GPQA-Diamond: struct<repository: string, revision: string, split: string>, HumanEval: struct< (... 136 chars omitted)
child 0, GPQA-Diamond: struct<repository: string, revision: string, split: string>
child 0, repository: string
child 1, revision: string
child 2, split: string
child 1, HumanEval: struct<repository: string, revision: string, split: string>
child 0, repository: string
child 1, revision: string
child 2, split: string
child 2, MMLU: struct<config: string, repository: string, revision: string, split: string>
child 0, config: string
child 1, repository: string
child 2, revision: string
child 3, split: string
benchmark_text_fields: struct<GPQA-Diamond: list<item: string>, HumanEval: list<item: string>, MMLU: list<item: string>>
child 0, GPQA-Diamond: list<item: string>
child 0, item: string
child 1, HumanEval: list<item: string>
child 0, item: string
child 2, MMLU: list<item: string>
child 0, item: string
contexts_scanned: int64
exact_normalized_text_overlaps: list<item: null>
child 0, item: null
overlap_count: int64
status: string
verified_payload_hashes: bool
sentinel_hidden_size_bytes: int64
allocation_counts: struct<chinese: int64, code_tests_documentation_issues: int64, dialogue_instruction_assistance: int6 (... 254 chars omitted)
child 0, chinese: int64
child 1, code_tests_documentation_issues: int64
child 2, dialogue_instruction_assistance: int64
child 3, encyclopedic_factual: int64
child 4, literary_narrative_creative: int64
child 5, news_history_economics_legal_essays: int64
child 6, other_multilingual: int64
child 7, scientific_technical: int64
child 8, structured_data_tools_apis_tables: int64
child 9, worked_math_science_formal_reasoning: int64
reference_hidden_size_bytes: int64
sentinel_repeat_mean_kld: struct<00-vs-01: double, 00-vs-02: double, 01-vs-02: double>
child 0, 00-vs-01: double
child 1, 00-vs-02: double
child 2, 01-vs-02: double
lm_head_size_bytes: int64
source_cluster_count: int64
context_count: int64
sentinel_count: int64
partition_counts: struct<analysis: int64, qualification: int64>
child 0, analysis: int64
child 1, qualification: int64
sentinel_live_logit_size_bytes: int64
to
{'allocation_counts': {'chinese': Value('int64'), 'code_tests_documentation_issues': Value('int64'), 'dialogue_instruction_assistance': Value('int64'), 'encyclopedic_factual': Value('int64'), 'literary_narrative_creative': Value('int64'), 'news_history_economics_legal_essays': Value('int64'), 'other_multilingual': Value('int64'), 'scientific_technical': Value('int64'), 'structured_data_tools_apis_tables': Value('int64'), 'worked_math_science_formal_reasoning': Value('int64')}, 'context_count': Value('int64'), 'lm_head_size_bytes': Value('int64'), 'partition_counts': {'analysis': Value('int64'), 'qualification': Value('int64')}, 'reference_hidden_size_bytes': Value('int64'), 'sentinel_count': Value('int64'), 'sentinel_hidden_size_bytes': Value('int64'), 'sentinel_live_logit_size_bytes': Value('int64'), 'sentinel_repeat_mean_kld': {'00-vs-01': Value('float64'), '00-vs-02': Value('float64'), '01-vs-02': Value('float64')}, 'source_cluster_count': Value('int64'), 'status': Value('string'), 'verified_payload_hashes': Value('bool')}
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
benchmark_revisions: struct<GPQA-Diamond: struct<repository: string, revision: string, split: string>, HumanEval: struct< (... 136 chars omitted)
child 0, GPQA-Diamond: struct<repository: string, revision: string, split: string>
child 0, repository: string
child 1, revision: string
child 2, split: string
child 1, HumanEval: struct<repository: string, revision: string, split: string>
child 0, repository: string
child 1, revision: string
child 2, split: string
child 2, MMLU: struct<config: string, repository: string, revision: string, split: string>
child 0, config: string
child 1, repository: string
child 2, revision: string
child 3, split: string
benchmark_text_fields: struct<GPQA-Diamond: list<item: string>, HumanEval: list<item: string>, MMLU: list<item: string>>
child 0, GPQA-Diamond: list<item: string>
child 0, item: string
child 1, HumanEval: list<item: string>
child 0, item: string
child 2, MMLU: list<item: string>
child 0, item: string
contexts_scanned: int64
exact_normalized_text_overlaps: list<item: null>
child 0, item: null
overlap_count: int64
status: string
verified_payload_hashes: bool
sentinel_hidden_size_bytes: int64
allocation_counts: struct<chinese: int64, code_tests_documentation_issues: int64, dialogue_instruction_assistance: int6 (... 254 chars omitted)
child 0, chinese: int64
child 1, code_tests_documentation_issues: int64
child 2, dialogue_instruction_assistance: int64
child 3, encyclopedic_factual: int64
child 4, literary_narrative_creative: int64
child 5, news_history_economics_legal_essays: int64
child 6, other_multilingual: int64
child 7, scientific_technical: int64
child 8, structured_data_tools_apis_tables: int64
child 9, worked_math_science_formal_reasoning: int64
reference_hidden_size_bytes: int64
sentinel_repeat_mean_kld: struct<00-vs-01: double, 00-vs-02: double, 01-vs-02: double>
child 0, 00-vs-01: double
child 1, 00-vs-02: double
child 2, 01-vs-02: double
lm_head_size_bytes: int64
source_cluster_count: int64
context_count: int64
sentinel_count: int64
partition_counts: struct<analysis: int64, qualification: int64>
child 0, analysis: int64
child 1, qualification: int64
sentinel_live_logit_size_bytes: int64
to
{'allocation_counts': {'chinese': Value('int64'), 'code_tests_documentation_issues': Value('int64'), 'dialogue_instruction_assistance': Value('int64'), 'encyclopedic_factual': Value('int64'), 'literary_narrative_creative': Value('int64'), 'news_history_economics_legal_essays': Value('int64'), 'other_multilingual': Value('int64'), 'scientific_technical': Value('int64'), 'structured_data_tools_apis_tables': Value('int64'), 'worked_math_science_formal_reasoning': Value('int64')}, 'context_count': Value('int64'), 'lm_head_size_bytes': Value('int64'), 'partition_counts': {'analysis': Value('int64'), 'qualification': Value('int64')}, 'reference_hidden_size_bytes': Value('int64'), 'sentinel_count': Value('int64'), 'sentinel_hidden_size_bytes': Value('int64'), 'sentinel_live_logit_size_bytes': Value('int64'), 'sentinel_repeat_mean_kld': {'00-vs-01': Value('float64'), '00-vs-02': Value('float64'), '01-vs-02': Value('float64')}, 'source_cluster_count': Value('int64'), 'status': Value('string'), 'verified_payload_hashes': Value('bool')}
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.
Kimi K3 distribution-fidelity reference
Status
The artifact is qualified for paired, teacher-forced comparison of Kimi K3 compressed checkpoints against the official MXFP4 checkpoint. It contains 1,024 distinct 2,048-token source contexts and 2,096,128 scored next-token positions.
The primary metric is:
KL(official MXFP4 distribution || candidate distribution)
The reference stores BF16 hidden states after final RMSNorm and before the
language-model head. lm-head/weight.safetensors reconstructs the complete
163,840-token distribution through the included deterministic comparator.
Immutable identifiers
- suite manifest SHA-256:
f3a79f7f28365d406a19a82cf210c25adf18974c4b9b607ab3754e9939f941cf - suite token SHA-256:
70cd72175fcb4574123b90a83e2b71bb460b8c2938cec9573645dc8c9a46a3bd - official checkpoint:
moonshotai/Kimi-K3revision2496450e92e425c886db095102a52a6682ca3970 - canonical LM-head file SHA-256:
c282a205aed4d19fbbe4ff907feff979c50b24723e8f552ddc141ef008c5de11 - capture runtime manifest SHA-256:
ec02e1633b045d9e7d525b41a02f20b2e1ab08adceabb4338a678b716039a783
Run sha256sum --check checksums.txt after downloading. Mixed source licenses
are recorded per source in sources.json and source-registry.json; no single
license replaces those source-specific terms.
Contents
manifest.json
suite-manifest.json
capture-runtime.json
sources.json
source-registry.json
partitions.json
tokens/
reference-hidden/
lm-head/
sentinel-hidden/
sentinel-live-logits/
capture-tools/
comparators/
validation/
checksums.txt
Each reference-hidden/hidden_NNNN.safetensors file contains BF16
hidden_states with shape [2047, 7168]. The shared LM head contains BF16
weight with shape [163840, 7168].
The 768-context analysis partition supports quantization design and parameter selection. The 256-context qualification partition supports evaluation only after codec parameters and acceptance thresholds are frozen. Sixty-four analysis contexts are runtime-repeat sentinels.
Capture a candidate
The candidate server must use the model code, TP16/DCP1 topology, B12X MLA,
Triton KDA prefill, BF16 activations and KV cache, batching configuration, and
canonical LM-head weight recorded in capture-runtime.json. Runtime logs must
prove the intended checkpoint, quantization path, and kernels.
capture-tools/serve-kimi-k3-fidelity-capture.sh requires
CAPTURE_VLLM_SOURCE, CAPTURE_OUTPUT, and a writable CAPTURE_JIT_CACHE.
CAPTURE_HF_CACHE defaults to /root/.cache/huggingface.
python capture-tools/capture-kimi-k3-hidden-suite.py \
--url http://127.0.0.1:8001/v1/completions \
--model Kimi-K3 \
--suite-dir . \
--capture-dir /mnt/kld/candidate-runtime/capture-hidden \
--output-dir /mnt/kld/candidate-hidden \
--runtime-manifest /mnt/kld/candidate-runtime.json \
--run-name candidate-checkpoint-identifier \
--context-filter analysis \
--delete-raw-chunks-after-finalize
The client submits stored token IDs directly and resumes from an atomically updated manifest. It does not retokenize source text.
Compare a candidate
python comparators/compare-kimi-k3-hidden-replay.py \
--reference-hidden-dir reference-hidden \
--candidate-hidden-dir /mnt/kld/candidate-hidden \
--lm-head lm-head/weight.safetensors \
--suite-manifest suite-manifest.json \
--context-filter analysis \
--verify-source-file-hashes \
--output /mnt/kld/candidate-analysis-kld.json
The comparator processes the complete vocabulary in bounded chunks and reports token statistics, top-1 agreement, allocation-stratified source-cluster bootstrap intervals, depth buckets, class estimates, and high-disagreement source identities.
Compare two candidate result receipts with:
python comparators/compare-kimi-k3-fidelity-receipts.py \
--candidate-a-report /mnt/kld/candidate-a-analysis-kld.json \
--candidate-b-report /mnt/kld/candidate-b-analysis-kld.json \
--candidate-a-label candidate-a \
--candidate-b-label candidate-b \
--output /mnt/kld/candidate-a-vs-candidate-b.json
Negative difference_a_minus_b values favor candidate A. The paired report
includes mean and median context differences, source-cluster win counts,
allocation-stratified confidence intervals, and the largest disagreements.
Numerical qualification and repeat variation
Offline LM-head replay reproduced the 32-context live-logit reference with mean
KLD 1.229325e-6, p99.9 KLD 0.000274, maximum KLD 0.00195, and top-1
agreement 0.999954. Changing the vocabulary chunk size changed the aggregate
result by 1.49e-9.
Three official-checkpoint captures over the same 64 sentinels produced pairwise
mean KLD values from 0.0031337795 to 0.0032166686, with top-1 agreement from
0.98326056 to 0.98348192. The 64 sentinels represent 57 source clusters;
hidden-state replay and live-logit comparison produced identical metrics for
all three repeat pairs. Candidate improvements comparable to that repeat
variation require repeated capture and paired confidence intervals.
Limitations
This artifact measures teacher-forced distribution fidelity at 2,048 input tokens. It does not establish benchmark correctness, free-running generation quality, multimodal capability, tool-execution correctness, long-context behavior, or absence of cumulative routing drift.
The complete reproduction procedure is documented in the Kimi K3 distribution-fidelity runbook.
- Downloads last month
- 603