Datasets:

wrldsuksgo2mars
/

cuteafd-fidelity

Tasks:

Text Generation

Languages:

English

Size:

10K<n<100K

License:

Dataset card Files Files and versions

xet

Community

Dataset Viewer

Duplicate

Split (1)

train

The dataset viewer is not available for this split.

Cannot load the dataset split (in streaming mode) to extract the first rows.

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
calibration: struct<baseline_metrics: struct<decode: list<item: struct<confident_top1: double, kl: double, top1:  (... 1056 chars omitted)
  child 0, baseline_metrics: struct<decode: list<item: struct<confident_top1: double, kl: double, top1: double, top3: double>>, p (... 91 chars omitted)
      child 0, decode: list<item: struct<confident_top1: double, kl: double, top1: double, top3: double>>
          child 0, item: struct<confident_top1: double, kl: double, top1: double, top3: double>
              child 0, confident_top1: double
              child 1, kl: double
              child 2, top1: double
              child 3, top3: double
      child 1, prefill: list<item: struct<confident_top1: double, kl: double, top1: double, top3: double>>
          child 0, item: struct<confident_top1: double, kl: double, top1: double, top3: double>
              child 0, confident_top1: double
              child 1, kl: double
              child 2, top1: double
              child 3, top3: double
  child 1, common_floor: struct<kl_max: double, top1_min: double>
      child 0, kl_max: double
      child 1, top1_min: double
  child 2, comparison_policy_commit: string
  child 3, comparison_sha256: string
  child 4, coordinator_feature_commit: string
  child 5, coordinator_sha256: string
  child 6, coordinator_source_diff_sha256: string
  child 7, daemon_identity: string
  child 8, expect_by_path: struct<decode: struct<kl_max: double, top1_min: double, tripwires: struct<confident_drop_margi
...
hild 11, seconds: double
      child 12, set_sha256: string
      child 13, snapshot_identity: struct<checkpoint: string, config_sha256: string, snapshot_revision: string, tokenizer_sha256: strin (... 2 chars omitted)
          child 0, checkpoint: string
          child 1, config_sha256: string
          child 2, snapshot_revision: string
          child 3, tokenizer_sha256: string
      child 14, source_window: string
      child 15, vocab: int64
  child 2, reference: string
  child 3, snapshot_identity: struct<checkpoint: string, config_sha256: string, snapshot_revision: string, tokenizer_sha256: strin (... 2 chars omitted)
      child 0, checkpoint: string
      child 1, config_sha256: string
      child 2, snapshot_revision: string
      child 3, tokenizer_sha256: string
reference_sha256: string
root_checkpoint: struct<checkpoint: string, config_sha256: string, id: string, precision: string, snapshot_revision:  (... 33 chars omitted)
  child 0, checkpoint: string
  child 1, config_sha256: string
  child 2, id: string
  child 3, precision: string
  child 4, snapshot_revision: string
  child 5, tokenizer_sha256: string
schema: string
set_sha256: string
set_version: string
source_reference_sha256: string
top_k: int64
vocab: int64
windows_sha256: string
configs: list<item: struct<name: string, path: string, sha256: string>>
  child 0, item: struct<name: string, path: string, sha256: string>
      child 0, name: string
      child 1, path: string
      child 2, sha256: string
to
{'schema': Value('string'), 'configs': List({'name': Value('string'), 'path': Value('string'), 'sha256': 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 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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
              calibration: struct<baseline_metrics: struct<decode: list<item: struct<confident_top1: double, kl: double, top1:  (... 1056 chars omitted)
                child 0, baseline_metrics: struct<decode: list<item: struct<confident_top1: double, kl: double, top1: double, top3: double>>, p (... 91 chars omitted)
                    child 0, decode: list<item: struct<confident_top1: double, kl: double, top1: double, top3: double>>
                        child 0, item: struct<confident_top1: double, kl: double, top1: double, top3: double>
                            child 0, confident_top1: double
                            child 1, kl: double
                            child 2, top1: double
                            child 3, top3: double
                    child 1, prefill: list<item: struct<confident_top1: double, kl: double, top1: double, top3: double>>
                        child 0, item: struct<confident_top1: double, kl: double, top1: double, top3: double>
                            child 0, confident_top1: double
                            child 1, kl: double
                            child 2, top1: double
                            child 3, top3: double
                child 1, common_floor: struct<kl_max: double, top1_min: double>
                    child 0, kl_max: double
                    child 1, top1_min: double
                child 2, comparison_policy_commit: string
                child 3, comparison_sha256: string
                child 4, coordinator_feature_commit: string
                child 5, coordinator_sha256: string
                child 6, coordinator_source_diff_sha256: string
                child 7, daemon_identity: string
                child 8, expect_by_path: struct<decode: struct<kl_max: double, top1_min: double, tripwires: struct<confident_drop_margi
              ...
              hild 11, seconds: double
                    child 12, set_sha256: string
                    child 13, snapshot_identity: struct<checkpoint: string, config_sha256: string, snapshot_revision: string, tokenizer_sha256: strin (... 2 chars omitted)
                        child 0, checkpoint: string
                        child 1, config_sha256: string
                        child 2, snapshot_revision: string
                        child 3, tokenizer_sha256: string
                    child 14, source_window: string
                    child 15, vocab: int64
                child 2, reference: string
                child 3, snapshot_identity: struct<checkpoint: string, config_sha256: string, snapshot_revision: string, tokenizer_sha256: strin (... 2 chars omitted)
                    child 0, checkpoint: string
                    child 1, config_sha256: string
                    child 2, snapshot_revision: string
                    child 3, tokenizer_sha256: string
              reference_sha256: string
              root_checkpoint: struct<checkpoint: string, config_sha256: string, id: string, precision: string, snapshot_revision:  (... 33 chars omitted)
                child 0, checkpoint: string
                child 1, config_sha256: string
                child 2, id: string
                child 3, precision: string
                child 4, snapshot_revision: string
                child 5, tokenizer_sha256: string
              schema: string
              set_sha256: string
              set_version: string
              source_reference_sha256: string
              top_k: int64
              vocab: int64
              windows_sha256: string
              configs: list<item: struct<name: string, path: string, sha256: string>>
                child 0, item: struct<name: string, path: string, sha256: string>
                    child 0, name: string
                    child 1, path: string
                    child 2, sha256: string
              to
              {'schema': Value('string'), 'configs': List({'name': Value('string'), 'path': Value('string'), 'sha256': Value('string')})}
              because column names don't match

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.

CuteAFD Fidelity

This is a numerical-fidelity panel for comparing inference precision arms, not a training corpus or a general model-quality ranking. Each config binds one model family, set version, official root checkpoint and qualified reference arithmetic. Additional model configs will be added over time.

Configs

Every config has 64 windows and 32,768 scored positions. Assistant-generated (gen) positions are distinct from conditioning or repository-source (ctx) positions; paired primary gates use gen only. Decode-shaped and prefill-shaped engine passes are both required for a full-tier precision verdict. Each config's generator settings, root checkpoint and calibrated expect values are in its manifest and qualification files.

Config	Checkpoint (root)	Revision	`gen` positions	Calibrated `expect`
`deepseek_v41-v2_20261005`	`deepseek-ai/DeepSeek-V4.1-Flash`, official FP8-block/MXFP4 (no vendor BF16 original is published)	`dba1be0a40aa45a94ad051997016db3960a90277`	17,656	top-1 ≥ 94%, KL ≤ 0.04
`mimo_v2-v2_20261005_flash_mopd` (superseded by `mimo_v2-v2_20261007_flash_mopd_qkvfixed`)	`XiaomiMiMo/MiMo-V2.6-Flash-MOPD`, checkpoint precision	`2479e2d0029eca9a34cc7e7f55a121925f81908e`	11,565	top-1 ≥ 92%, KL ≤ 0.06; confide
nt top-1 ≥ 95%, top-3 ≥ 97%
`glm5_flash-v2_20261005_bf16root`	`zai-org/GLM-5.3-Flash-BF16`, vendor BF16 original (text generated by the official FP8 release)	`a5b45eb41df6402735dedc900be14a42e8d5e538`	15,104	top-1 ≥ 93%, KL ≤ 0.05; confident top-1 ≥ 96%, top-3 ≥ 97%
`qwen4-v2_20261005_fp8`	`Qwen/Qwen3.8-Flash-Next`, vendor BF16 original (text generated by the official FP8 release)	`de4b8e4d43b917e7706784d8bb445c9af86a3540`	16,253	top-1 ≥ 94%, KL ≤ 0.04; confident top-1 ≥ 96%, top-3 ≥ 97%
`deepseek_v4-v2_20261005_v4flash`	`deepseek-ai/DeepSeek-V4-Flash-0731`, official FP8/FP4 release (no vendor BF16 original is published)	`9e165c30e2704aec5d9d593cce3eebd58bbef1cb`	17,401	top-1 ≥ 94%, KL ≤ 0.04; confident top-1 ≥ 96%, top-3 ≥ 97%
`glm5-v2_20261006_exl3k4_bf16root`	`zai-org/GLM-5.3-BF16`, vendor BF16 original (text generated by `wrldsuksgo2mars/GLM-5.3-EXL3-K4-v1`, TP4)	`9d2398f478cab2de883137db3a36ad2c96205e24`	15,459	top-1 ≥ 91%, KL ≤ 0.06; confident top-1 ≥ 94%, top-3 ≥ 97%
`deepseek_v4-v2_20261005_v4pro`	`deepseek-ai/DeepSeek-V4-Pro-0813`, official FP8 checkpoint (text generated by `wrldsuksgo2mars/DeepSeek-V4-Pro-0813-EXL3-K2-calibrated-v1`, TP4)	`72e1d3230f6c080a530b0a1d46f8eb4602340597`	18,054	top-1 ≥ 91%, KL ≤ 0.06; confident top-1 ≥ 95%, top-3 ≥ 97%

Superseded: mimo_v2-v2_20261005_flash_mopd. Its reference log-probabilities were computed with a fused-QKV FP8 dequantization error (MiMo V2.6 stores one 128-row scale grid per tensor-parallel shard over [q | k | v]; the reference read one grid per segment), and its text was generated by an engine with the same error. Use mimo_v2-v2_20261007_flash_mopd_qkvfixed: the engine, the reference and the text set were all regenerated with the fix. Found by Hugh Madden (cuteafd issue #3).

Three policies are separate:

a common absolute floor: top-1 ≥ 90%, KL ≤ 0.06 nat;
each config's expect, calibrated from its own repeated baselines;
the paired precision bar: top-1 loss < 0.005 and KL increase < 0.005 nat, both as one-sided 95% bounds on both full shapes.

Each config after V4.1 has its own README.md in its folder, with its repeated baselines and tripwires.

Layout

configs.json: config names and manifest checksums.
<config>/manifest.json: set and reference hashes, root checkpoint/precision, snapshot/config/tokenizer hashes, generator and reference provenance, window file hashes, source full-row hashes and qualification checksum.
<config>/windows.json: exact token ids, role mask, score start, bucket and source provenance. No user conversations or external teacher dataset are used. Deployment metadata is sanitized: native artifact SHA256/fork revision and expert topology replace local paths and addresses. The original sealed set identity is retained; publication_provenance declares this metadata-only projection. Scored tokens, roles and geometry are unchanged.
<config>/<window>.safetensors: top_ids U32 [N,1024], top_log_probs F16 [N,1024], tail_log_mass F32 [N], next_token_log_prob F32 [N], positions U32 [N], next_token_ids U32 [N], roles U8 [N] (1=gen, 0=ctx).
<config>/qualification.json: paired top-1024+tail versus full-vocabulary validation for decode and prefill.

Top support is reference-fixed, descending by reference log-probability with ascending token id as the tie break. Top values retain the original sealed f16 log-probabilities. Tail log-mass is the log-sum-exp of all omitted sealed reference values, rounded to f32. Normalize the 1024 entries and tail together before KL; f16 rounding changes their total mass. Next-token log-probability is stored separately, normalized against the full reference vocabulary, including when the next token is outside the top support.

Score the engine on the same 1024 ids plus its aggregate remaining probability mass. This coarse-grained KL is a lower bound on full-vocabulary KL, not a claim that their absolute values are identical. Do not replace the engine tail with a reference tail or compare different supports. Keep top-1/NLL/tripwire gates and window-block bootstrap provenance unchanged.

Usage

Use a cuteafd build with the dataset loader. Replace the revision placeholder with the dataset's immutable 40-character Hugging Face commit, never main. The server must serve the checkpoint named by the config. Run each precision arm twice, once per scoring shape, with distinct new dump/output paths:

cuteafd bench fidelity run --tier full --arm baseline \
  --dataset wrldsuksgo2mars/cuteafd-fidelity \
  --dataset-config deepseek_v41-v2_20261005 \
  --dataset-revision <IMMUTABLE_HF_COMMIT> \
  --score-path decode --dump-dir /server-local-nvme/baseline-decode \
  --out baseline-decode.json

The client fetches revision-pinned files, verifies manifest/index checksums, normalizes the stored reference support and tail jointly, and records the repository/config/revision in its run result. Engine full-row dumps must be visible to the client; they are streamed per row to compute the engine's tail without cancellation from rounded top probabilities. Repeat with --score-path prefill and the candidate arm, then use cuteafd bench fidelity compare-full --a-decode candidate-decode.json --b-decode baseline-decode.json --a-prefill candidate-prefill.json --b-prefill baseline-prefill.json. Independent agentic replay remains a separate gate. Dataset qualification does not waive any statistical tripwire.

Qualification

On the saved V4.1 FP8-head paired comparison (17,656 gen positions, 53 windows; 5,000 window-bootstrap replicates, seed 20260829), compact paired KL deltas and one-sided 95% bounds agree with full-vocabulary values to at most 1.18e-7 nat, well inside the 1e-4 nat criterion. Both decode and prefill remain PASS. Full-row score reconstruction error is at most 1.11e-11 nat. Qualification includes f16 entry storage, f32 tail storage and joint normalization. This is evidence for this checkpoint/set/pair, not a guarantee for all future models.

MiMo Flash: two default-precision baseline repeats per full shape reproduce exactly (decode 94.0078% / 0.040078 nat, prefill 94.3450% / 0.039919 nat). Compact and full-vocabulary paired bounds agree, and full-KL reproduction error is at most 2.23e-11 nat. Details are in its qualification.json.

GLM 5.3 Flash: two default-precision baseline repeats per full shape reproduce exactly (decode 95.9613% / 0.019741 nat, prefill 95.7693% / 0.021644 nat). Compact and full-vocabulary paired bounds agree, and full-KL reproduction error is at most 2.54e-11 nat. F16 storage moves one generated agreement per arm and shape (at most 6.6e-5 absolute top-1), identically in both repeats.

Qwen 3.8 Flash Next: two default-precision baseline repeats per full shape reproduce exactly (decode 96.4622% / 0.011675 nat, prefill 96.3268% / 0.011692 nat). Compact and full-vocabulary paired bounds agree; full-KL reproduction error is at most 1.97e-11 nat. F16 storage moves one generated top-3 containment (6.2e-5 absolute), identically in both repeats.

DeepSeek V4 Flash: baseline repeats per full shape are statistically repeatable (paired upper95 top-1 loss ≤ 0.0014, KL increase ≤ 0.0010 nat) but not byte-identical: identical-config decode launches differed by 0.39 points of top-1. Compact and full-vocabulary bounds agree to 3.9e-8; full-KL reproduction error is at most 6.7e-12 nat. See its README for the open launch-variation finding.

GLM 5.3: two default-precision baseline repeats per full shape reproduce exactly (decode 93.7900% / 0.057536 nat, prefill 93.7512% / 0.059289 nat). Compact and full-vocabulary paired bounds agree. Windows d03, d04 and a25 carry large context-only disagreement and are retained unfiltered; see the config's README for the measured cross-architecture reference sensitivity on two rows.

DeepSeek V4 Pro: two default-precision baseline repeats per full shape reproduce exactly across all 32,768 scored records (decode 93.1483% / 0.047423 nat, prefill 93.2979% / 0.046984 nat). Compact and full-vocabulary paired bounds agree; full-score reproduction error is at most 1.45e-11 nat.

MiMo V2.6 Flash (QKV-fixed): two default-precision baseline repeats per full shape reproduce exactly (decode 95.7307% / 0.018233 nat, prefill 95.8449% / 0.019411 nat). Compact and full-vocabulary paired bounds agree; full-score reproduction error is at most 3.8e-11 nat. Against the superseded config (94.0% / 0.040 nat), the fix halves KL.

Licence and provenance

This dataset is published under the cuteafd project's MIT licence. The pinned checkpoints are identified (name and revision) only as provenance for the reference log-probabilities; no weights are included. Original fixture prompts and generated continuations are supplied by this project. All ten repository-source windows were audited against their exact historical SHA-sealed files and reproduced with the pinned checkpoint tokenizer: they contain only cuteafd's own MIT source, not third-party file text. The paths and SHA256 values are retained in windows.json. LICENSE is the cuteafd MIT licence and covers this dataset. Reference log-probabilities derive from the named official checkpoint; no weights are redistributed. Source hashes and reference-generation provenance are retained per config.

The panel does not reuse external teacher logits. The repository fixture material is deliberately synthetic/non-private. Validate every checksum before a run and record the pinned dataset revision and manifest reference hash in run results.

Downloads last month: 93

Total file size:

1.66 GB