Datasets:

malaiwah
/

GLM-5.3-Flash-fidelity-suite-v1

Tags:

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:    ValueError
Message:      Expected object or value
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
                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 number in row 0
              
              During handling of the above exception, another exception occurred:
              
              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 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

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.

GLM-5.3-Flash Fidelity Suite v1

Historical distribution-fidelity evidence for GLM-5.3-Flash (released 2026-08-26): BF16-reference and FP8-as-served hidden-state captures, a shared LM head, and receipts from the declared capture/replay path. Compatible candidate captures can be compared on matching published positions without holding the 643 GB reference; this is not a universal native-serving or task-quality score. Protocol: the Qwen3.8-27B fidelity-suite v5 methodology (hidden-state replay through one shared BF16 head, exact two-pass full-vocab KL).

Publication correction — 2026-09-08

This card now distinguishes the historical full run from the public capture subset and withdraws unsupported cross-lane ratios, replay-equivalence and noise-bound claims. No measured value, sealed receipt, capture, head or digest has changed. Old reports retain their original bytes and historical prose.

Public scope: each of reference-bf16-shard0/ and as-served-fp8-shard0/ contains 512 contexts, not the full 5,120-context run below. Read each capture-manifest-shard.json (complete: false) and capture-cut-point.json; the original capture-manifest-full.json records the historical full run, not current local availability. These captures are post-final-RMSNorm: apply the head only, not final norm again. Public shard bytes alone cannot rederive the full-run headline. Candidate captures are a separate prerequisite.

Historical full-run FP8 vs BF16 replay measurements

metric	value
token mean KLD	0.028104 nats
context macro mean (historical nominal 95% bootstrap interval)	0.028104 [0.027205, 0.028982]
median / p99 / p999 KLD	4.93e-03 / 3.54e-01 / 1.37e+00
top-1 agreement	0.9427
mean JSD (bits)	0.009201
scored positions	10,480,640 (5,120 contexts x 2,047)

The interval above is retained from reports/report-fp8-vs-bf16.json, which records 837 bootstrap clusters for 5,120 contexts. That count alone does not establish independent-document sampling or population coverage. These are fixed-corpus descriptive statistics, not deployment-population inference. The historical Flash final25 panel discussed later is a different panel.

Per-stratum mean KLD:

stratum	contexts	mean KLD
code	1024	0.025320
encyclopedic	1024	0.022285
literary	1024	0.032324
multilingual	1024	0.025154
scientific	1024	0.035436

Restricted scored-position geometry (positions 1024+ only): token mean KLD 0.018794, top-1 0.9512. Matching geometry alone does not establish llama.cpp comparability; reference, tokens, vocabulary, scope and lane must also match.

Receipts

BF16 live-vs-replayed: mean KLD 1.49e-02, top-1 0.95695 over 8 contexts (reports/qualify-bf16.json). This measured discrepancy is not equality or proof that replay reproduces native-served logits.
bf16 determinism: 20/32 sentinel contexts byte-identical across independent engine loads.
fp8 determinism: 29/32 sentinel contexts byte-identical across independent engine loads.
Head equality: FP8 repo lm_head/final-norm byte-identical to BF16: True/True. This supports the declared shared-head path, not equal hidden states, logits or padded probability mass; shared-head replay cannot assess a different candidate head's quantization error.

Known issue: run-to-run nondeterminism (first report)

The day-one vLLM measurements observed cross-launch differences: 12/32 BF16 sentinel contexts differed bytewise in the reported pair, with mean replay KL 8.7e-4 / top-1 0.9946 (reports/determinism-bf16.json and reports/determinism-noise-bf16.json). Within-launch repeatability was observed under that configuration, not established universally. Live-vs-replay on a separate launch measured ~1.5e-2 over 8 contexts (reports/qualify-bf16.json). Neither value is a universal noise floor, additive error term or upper bound.

The initial Triton-autotune root-cause attribution and the proposed cache-only "deterministic rerun" recipe are superseded by the intervention results below. Kernel/collective differences were investigated, but the evidence does not uniquely attribute the discrepancy or prove deterministic operation. Historical discussion: https://github.com/vllm-project/vllm/pull/53906#issuecomment-5433635837 . A shared replay path does not cancel all capture or serving differences; additional perturbations may amplify or cancel divergence.

Cross-pipeline comparison: vs brandonmusic/GLM-5.3-Flash-BF16-Teacher-Logits (separate pipeline, full-vocab fp32 logits): mean KLD(theirs||ours) 1.27e-02, top-1 0.96653 over 51,175 positions. This is a measured discrepancy, not independent reproduction of identical outputs.

Pins

what	value
BF16 reference	`zai-org/GLM-5.3-Flash-BF16` @ `b1967181a3917ae70a437f4884748f6b8e3a1f4d`
FP8 as-served	`zai-org/GLM-5.3-Flash` @ `3f1971b7b5f7a528c9c4ef6212c8785298a8c24a`
engine	vLLM glm53-flash docker image (digest in `reports/image-pin.txt`), TP8 H200, eager, TF32 off, BF16 KV
suite	5,120 ctx x 2,048 tok, held-out v5-lineage corpus, GLM tokenizer, 0 calibration-contamination hits
contamination boundary	exllamav3 standard_cal_data @ 0c49587a

suite/ - tokens + manifest (partitions: analysis/qualification/sentinels)
reference-bf16-shard0/, as-served-fp8-shard0/ - 512 contexts x [2047, 4096] bf16 hidden states each
head/ - shared BF16 lm_head (154,880 x 4,096) + final norm + extraction receipt
reports/ - the KLD reports and every receipt above; SHA256SUMS covers all files

Score your own quant

Capture final-norm hidden states of your quant over suite/tokens/ (teacher-forced, one context per forward), then replay against reference-bf16-shard0 through head/head.safetensors with the fidelity harness (tools published alongside; see malaiwah's qwen38-27b-fidelity-suite-v5 for the protocol paper trail).

Produced autonomously on rented 8x H200; contact: malaiwah.

Determinism interventions (receipts in reports/)

configuration	sentinel pairs byte-identical
unpinned (3 independent pairs)	20/32, 25/32, 20/32
+ single-config Triton autotune shim (verified active)	20/32
+ VLLM_ALLREDUCE_USE_SYMM_MEM=0, NCCL Ring/Simple/1ch, CUBLAS_WORKSPACE_CONFIG	31/32
+ VLLM_USE_DEEP_GEMM=0	28/32

The stack-pinned pair had the highest observed match count, but these few pairs do not uniquely identify a cause or establish a general flip rate. The single-config Triton intervention did not eliminate differences. Residual sources remain unidentified. FP8-side receipts record live-vs-replay 2.51e-2 / top-1 0.9414 over 8 contexts; the sentinel discrepancy is separately in reports/determinism-noise-fp8.json, not a universal noise bound.

Calibration-clean scope (added 2026-08-29)

brandonmusic proposed a community protocol for quantization-fidelity measurement and ran a 13-gram calibration-overlap scan of his sealed 25-window panel against its own calibration-role windows. One whole domain of the final windows shares 37-39% of its 13-grams with calibration material, despite the panel being clean at the document-hash level. Document-hash dedup is not enough. He excluded that domain, leaving a 17-window calibration-clean scope.

Every malaiwah number published on that panel used all 25 windows, so it carries the same contamination. Recomputed on his clean scope from our own published per-window arrays - no GPU, no re-measurement, arithmetic on already-public data:

row	panel25	clean17	move
K6 sealed	0.013723	0.011677	-14.91 %
K8	0.012384	0.010829	-12.55 %
official FP8	0.020615	0.018665	-9.46 %
BF16 floor (cross-stack)	0.012712	0.010648	-16.24 %
brandonmusic 4bpw	0.024555	0.024949	+1.61 %

Claim correction, 2026-09-08: panel25 values remain unchanged, but the former 1.50x -> 1.60x K6/FP8 headline mixed the sealed-ep8 checkpoint lane with the cross-stack FP8 lane. It is withdrawn as a quality ratio. K8 is streaming; brandonmusic's row is his separate stack. This table is a mixed-design inventory, not a ranking. The reported 1.44% increase in FP8 excess over its cross-stack control does not establish that subtraction is generally stable or causal. Never substitute that control for the streaming control or mix panel scopes.

The final25 panel contains four source documents; clean17 contains three. Window-level intervals and signs are not independent population evidence. The report's historical paired contrasts also mix lanes; consult the corrected measurement semantics and model cards before using any contrast. Published means and report bytes stay unchanged.

reports/clean-scope-recompute.json - full recompute: both scopes, BCa intervals, per-domain tables, paired comparisons, provenance. sha256 822049d44d307631046b7a98a4ed6cbc223b7dd242d03be57f71436585a15780
reports/clean-scope-recompute.md - the same, rendered as tables. sha256 e6fbdc75c75ca6ecba95d4c934f38043cc38b4b92f1c95e97f3747b692a3bcc5

Both are emitted by bin/emit_clean_scope_report.py from committed per-window data, so anyone can regenerate them. Hashes are given here because the repo-root SHA256SUMS is the original v1-publish snapshot and does not cover reports added after it.

The scan, the threshold and the finding are brandonmusic's. Working: PROTOCOL-ALIGNMENT.md.

Related / follow-on work

Published K6/K8 artifacts and their scoped fidelity evidence: K6, K8. Mixed-rate assembly needs its own measurement; shared-seed parts do not prove a fresh mixed-rate encode or native-forward equivalence.
Native exllamav3 glm5_next port design: port/.
Calibration activations dataset: malaiwah/GLM-5.3-Flash-calibration-activations-v1.