The dataset viewer is not available for this split.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
cells: list<item: struct<cell_id: string, group: string, arm: string, context_tokens: int64, n: int64, corr (... 59 chars omitted)
child 0, item: struct<cell_id: string, group: string, arm: string, context_tokens: int64, n: int64, correct: int64, (... 47 chars omitted)
child 0, cell_id: string
child 1, group: string
child 2, arm: string
child 3, context_tokens: int64
child 4, n: int64
child 5, correct: int64
child 6, generated_recall: bool
child 7, qualification: string
hash_scheme: string
prompt_suffix_template: string
diagnostic: string
LSA_CITATION.md: string
manifests/rtx3090_historical_131072_graft_dense.json: string
scores/rtx3090_historical_131072_graft_qsa.jsonl: string
manifests/rtx3090_historical_261120_dense.json: string
manifests/rtx3090_historical_261120_b16.json: string
scores.csv: string
scores/rtx3090_historical_131072_graft_dense.jsonl: string
scores/rtx3090ti_reference_mask_196608_native_qsa.jsonl: string
manifests/rtx3090ti_196608_b128.json: string
scores/rtx3090ti_196608_b128.jsonl: string
scores/rtx3090_historical_131072_native_dense.jsonl: string
scores/rtx3090_historical_261120_dense.jsonl: string
scores/rtx3090_historical_261120_b128.jsonl: string
manifests/rtx3090ti_261120_b128.json: string
inputs/131072.json: string
manifests/rtx3090ti_reference_mask_196608_native_dense.json: string
manifests/rtx3090_historical_131072_native_dense.json: string
inputs/196608.json: string
manifests/rtx3090_historical_131072_native_qsa.json: string
manifests/rtx3090_historical_196608_dense.json: string
scores/rtx3090ti_261120_b128.jsonl: string
index.json: string
manifests/rtx3090ti_reference_mask_196608_native_qsa.json: string
scores/rtx3090ti_261120_dense.jsonl: string
scores/rtx3090_historical_131072_native_qsa.jsonl: string
scores/rtx3090ti_196608_dense.jsonl: string
README.md: string
scores/rtx3090_historical_261120_b16.jsonl: string
scores/rtx3090ti_reference_mask_196608_native_dense.jsonl: string
scores/rtx3090ti_196608_b16.jsonl: string
inputs/261120.json: string
manifests/rtx3090_historical_131072_dense.json: string
manifests/rtx3090_historical_131072_graft_qsa.json: string
manifests/rtx3090_historical_131072_b16.json: string
manifests/rtx3090_historical_196608_b16.json: string
manifests/rtx3090_historical_131072_b128.json: string
manifests/rtx3090ti_196608_dense.json: string
manifests/rtx3090_historical_261120_b128.json: string
manifests/rtx3090ti_196608_b16.json: string
manifests/rtx3090ti_261120_dense.json: string
scores/rtx3090_historical_196608_b16.jsonl: string
scores/rtx3090_historical_196608_dense.jsonl: string
scores/rtx3090_historical_196608_b128.jsonl: string
manifests/rtx3090_historical_196608_b128.json: string
scores/rtx3090_historical_131072_b16.jsonl: string
manifests/rtx3090ti_261120_b16.json: string
scores/rtx3090_historical_131072_b128.jsonl: string
scores/rtx3090_historical_131072_dense.jsonl: string
scores/rtx3090ti_261120_b16.jsonl: string
to
{'README.md': Value('string'), 'LSA_CITATION.md': Value('string'), 'index.json': Value('string'), 'scores.csv': Value('string'), 'inputs/261120.json': Value('string'), 'inputs/196608.json': Value('string'), 'inputs/131072.json': Value('string'), 'scores/rtx3090_historical_131072_b128.jsonl': Value('string'), 'scores/rtx3090_historical_131072_dense.jsonl': Value('string'), 'scores/rtx3090_historical_131072_graft_qsa.jsonl': Value('string'), 'scores/rtx3090_historical_261120_b16.jsonl': Value('string'), 'scores/rtx3090_historical_196608_dense.jsonl': Value('string'), 'scores/rtx3090_historical_196608_b128.jsonl': Value('string'), 'scores/rtx3090ti_196608_dense.jsonl': Value('string'), 'scores/rtx3090_historical_261120_dense.jsonl': Value('string'), 'scores/rtx3090_historical_131072_native_dense.jsonl': Value('string'), 'scores/rtx3090_historical_131072_native_qsa.jsonl': Value('string'), 'scores/rtx3090_historical_131072_b16.jsonl': Value('string'), 'scores/rtx3090ti_196608_b128.jsonl': Value('string'), 'scores/rtx3090_historical_261120_b128.jsonl': Value('string'), 'scores/rtx3090ti_261120_b16.jsonl': Value('string'), 'scores/rtx3090ti_reference_mask_196608_native_qsa.jsonl': Value('string'), 'scores/rtx3090ti_261120_b128.jsonl': Value('string'), 'scores/rtx3090_historical_196608_b16.jsonl': Value('string'), 'scores/rtx3090ti_196608_b16.jsonl': Value('string'), 'scores/rtx3090ti_reference_mask_196608_native_dense.jsonl': Value('string'), 'scores/rtx3090_historical_131072_graft_dense.jsonl': Value('string'), 'scores/rtx3090ti_261120_dense.jsonl': Value('string'), 'manifests/rtx3090ti_196608_b128.json': Value('string'), 'manifests/rtx3090_historical_131072_b128.json': Value('string'), 'manifests/rtx3090_historical_196608_b16.json': Value('string'), 'manifests/rtx3090ti_reference_mask_196608_native_qsa.json': Value('string'), 'manifests/rtx3090_historical_261120_b128.json': Value('string'), 'manifests/rtx3090ti_261120_dense.json': Value('string'), 'manifests/rtx3090ti_reference_mask_196608_native_dense.json': Value('string'), 'manifests/rtx3090_historical_131072_dense.json': Value('string'), 'manifests/rtx3090_historical_196608_b128.json': Value('string'), 'manifests/rtx3090_historical_131072_b16.json': Value('string'), 'manifests/rtx3090_historical_261120_dense.json': Value('string'), 'manifests/rtx3090ti_196608_dense.json': Value('string'), 'manifests/rtx3090ti_261120_b16.json': Value('string'), 'manifests/rtx3090ti_261120_b128.json': Value('string'), 'manifests/rtx3090_historical_261120_b16.json': Value('string'), 'manifests/rtx3090_historical_131072_native_qsa.json': Value('string'), 'manifests/rtx3090_historical_131072_graft_qsa.json': Value('string'), 'manifests/rtx3090ti_196608_b16.json': Value('string'), 'manifests/rtx3090_historical_131072_graft_dense.json': Value('string'), 'manifests/rtx3090_historical_196608_dense.json': Value('string'), 'manifests/rtx3090_historical_131072_native_dense.json': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, 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 129, 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 544, 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 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, 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 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, 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 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
cells: list<item: struct<cell_id: string, group: string, arm: string, context_tokens: int64, n: int64, corr (... 59 chars omitted)
child 0, item: struct<cell_id: string, group: string, arm: string, context_tokens: int64, n: int64, correct: int64, (... 47 chars omitted)
child 0, cell_id: string
child 1, group: string
child 2, arm: string
child 3, context_tokens: int64
child 4, n: int64
child 5, correct: int64
child 6, generated_recall: bool
child 7, qualification: string
hash_scheme: string
prompt_suffix_template: string
diagnostic: string
LSA_CITATION.md: string
manifests/rtx3090_historical_131072_graft_dense.json: string
scores/rtx3090_historical_131072_graft_qsa.jsonl: string
manifests/rtx3090_historical_261120_dense.json: string
manifests/rtx3090_historical_261120_b16.json: string
scores.csv: string
scores/rtx3090_historical_131072_graft_dense.jsonl: string
scores/rtx3090ti_reference_mask_196608_native_qsa.jsonl: string
manifests/rtx3090ti_196608_b128.json: string
scores/rtx3090ti_196608_b128.jsonl: string
scores/rtx3090_historical_131072_native_dense.jsonl: string
scores/rtx3090_historical_261120_dense.jsonl: string
scores/rtx3090_historical_261120_b128.jsonl: string
manifests/rtx3090ti_261120_b128.json: string
inputs/131072.json: string
manifests/rtx3090ti_reference_mask_196608_native_dense.json: string
manifests/rtx3090_historical_131072_native_dense.json: string
inputs/196608.json: string
manifests/rtx3090_historical_131072_native_qsa.json: string
manifests/rtx3090_historical_196608_dense.json: string
scores/rtx3090ti_261120_b128.jsonl: string
index.json: string
manifests/rtx3090ti_reference_mask_196608_native_qsa.json: string
scores/rtx3090ti_261120_dense.jsonl: string
scores/rtx3090_historical_131072_native_qsa.jsonl: string
scores/rtx3090ti_196608_dense.jsonl: string
README.md: string
scores/rtx3090_historical_261120_b16.jsonl: string
scores/rtx3090ti_reference_mask_196608_native_dense.jsonl: string
scores/rtx3090ti_196608_b16.jsonl: string
inputs/261120.json: string
manifests/rtx3090_historical_131072_dense.json: string
manifests/rtx3090_historical_131072_graft_qsa.json: string
manifests/rtx3090_historical_131072_b16.json: string
manifests/rtx3090_historical_196608_b16.json: string
manifests/rtx3090_historical_131072_b128.json: string
manifests/rtx3090ti_196608_dense.json: string
manifests/rtx3090_historical_261120_b128.json: string
manifests/rtx3090ti_196608_b16.json: string
manifests/rtx3090ti_261120_dense.json: string
scores/rtx3090_historical_196608_b16.jsonl: string
scores/rtx3090_historical_196608_dense.jsonl: string
scores/rtx3090_historical_196608_b128.jsonl: string
manifests/rtx3090_historical_196608_b128.json: string
scores/rtx3090_historical_131072_b16.jsonl: string
manifests/rtx3090ti_261120_b16.json: string
scores/rtx3090_historical_131072_b128.jsonl: string
scores/rtx3090_historical_131072_dense.jsonl: string
scores/rtx3090ti_261120_b16.jsonl: string
to
{'README.md': Value('string'), 'LSA_CITATION.md': Value('string'), 'index.json': Value('string'), 'scores.csv': Value('string'), 'inputs/261120.json': Value('string'), 'inputs/196608.json': Value('string'), 'inputs/131072.json': Value('string'), 'scores/rtx3090_historical_131072_b128.jsonl': Value('string'), 'scores/rtx3090_historical_131072_dense.jsonl': Value('string'), 'scores/rtx3090_historical_131072_graft_qsa.jsonl': Value('string'), 'scores/rtx3090_historical_261120_b16.jsonl': Value('string'), 'scores/rtx3090_historical_196608_dense.jsonl': Value('string'), 'scores/rtx3090_historical_196608_b128.jsonl': Value('string'), 'scores/rtx3090ti_196608_dense.jsonl': Value('string'), 'scores/rtx3090_historical_261120_dense.jsonl': Value('string'), 'scores/rtx3090_historical_131072_native_dense.jsonl': Value('string'), 'scores/rtx3090_historical_131072_native_qsa.jsonl': Value('string'), 'scores/rtx3090_historical_131072_b16.jsonl': Value('string'), 'scores/rtx3090ti_196608_b128.jsonl': Value('string'), 'scores/rtx3090_historical_261120_b128.jsonl': Value('string'), 'scores/rtx3090ti_261120_b16.jsonl': Value('string'), 'scores/rtx3090ti_reference_mask_196608_native_qsa.jsonl': Value('string'), 'scores/rtx3090ti_261120_b128.jsonl': Value('string'), 'scores/rtx3090_historical_196608_b16.jsonl': Value('string'), 'scores/rtx3090ti_196608_b16.jsonl': Value('string'), 'scores/rtx3090ti_reference_mask_196608_native_dense.jsonl': Value('string'), 'scores/rtx3090_historical_131072_graft_dense.jsonl': Value('string'), 'scores/rtx3090ti_261120_dense.jsonl': Value('string'), 'manifests/rtx3090ti_196608_b128.json': Value('string'), 'manifests/rtx3090_historical_131072_b128.json': Value('string'), 'manifests/rtx3090_historical_196608_b16.json': Value('string'), 'manifests/rtx3090ti_reference_mask_196608_native_qsa.json': Value('string'), 'manifests/rtx3090_historical_261120_b128.json': Value('string'), 'manifests/rtx3090ti_261120_dense.json': Value('string'), 'manifests/rtx3090ti_reference_mask_196608_native_dense.json': Value('string'), 'manifests/rtx3090_historical_131072_dense.json': Value('string'), 'manifests/rtx3090_historical_196608_b128.json': Value('string'), 'manifests/rtx3090_historical_131072_b16.json': Value('string'), 'manifests/rtx3090_historical_261120_dense.json': Value('string'), 'manifests/rtx3090ti_196608_dense.json': Value('string'), 'manifests/rtx3090ti_261120_b16.json': Value('string'), 'manifests/rtx3090ti_261120_b128.json': Value('string'), 'manifests/rtx3090_historical_261120_b16.json': Value('string'), 'manifests/rtx3090_historical_131072_native_qsa.json': Value('string'), 'manifests/rtx3090_historical_131072_graft_qsa.json': Value('string'), 'manifests/rtx3090ti_196608_b16.json': Value('string'), 'manifests/rtx3090_historical_131072_graft_dense.json': Value('string'), 'manifests/rtx3090_historical_196608_dense.json': Value('string'), 'manifests/rtx3090_historical_131072_native_dense.json': Value('string')}
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.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Long-context recall with LSA and QSA
The 192K/255K same-card dense and LSA reruns are complete. The QSA rows are pending, will be added, and the native QSA default cached-selector replay issue is unresolved. The historical 128K table below must not be presented as a validated comparison with the published default QSA path.
LSA attribution: paper citation and qualified quality summary. Data and per-arm manifests: jakeatx/long-context-lsa-qsa-evaluation.
This is a fresh synthetic test, not a reproduction of QSA's published question set. We use the same metric, the share of answers fully correct, with 70 questions at each depth: seven task families at ten record positions. The document contains source-code filler, 15% unrelated look-alike filler, and one-character key decoys. The public recall question set and scorer were not available in the inspected pinned sources. We did not reproduce the separate published attention-mass capture benchmark.
LSA-128 and LSA-16 use qwen3.8-27b-Swift1.5-Uncensored-ATX-AI4.gguf, SJ-KVaRN KV, adaptive MTP with mb-MTP on, temperature 1, top-k 20, top-p 0.95, min-p 0, and seed 761115. Sampler order is penalties, dry, top-n-sigma, top-k, typical-p, top-p, min-p, XTC, temperature; penalties and optional filters are neutral. The dense reference uses the same checkpoint and settings with LSA off. LSA retains a nominal 32,896 historical tokens: 128 sinks, an 8,192-token window, and 24,576 selected tokens. Actual deduplicated and per-layer totals are not established by that nominal sum.
Native QSA uses the published fork at 864934c, rebuilt for sm86, and UD-Q4_K_M weights, Q8 KV, K2048, greedy decoding, MTP off. The QSA port uses its own grafted checkpoint and dense reference. These model, sampler, and attention-budget differences prevent an equal-budget or code-only attribution from this recall table. Probability comparisons use each implementation's own dense reference.
Historical 128K screen
All five answer arms below ran on one RTX 3090 at 350 W, PCIe gen3 x16, AMD EPYC 7551P. Each score has 70 questions. The native QSA cell used full-state prompt replay; later controls found disagreement between the default live cached selector and restored/reference paths. Thus the native row describes the recorded replay configuration, not established default-path behavior. The port row has not undergone the newer document-replay controls. Other cards' protocol pilots are not pooled here.
| Tokens | Arm | Fully correct | Descriptive 95% CI | First-token KLD vs own dense | Same first token |
|---|---|---|---|---|---|
| 131,072 | Dense | 63/70 (90.0%) | 80.77–95.07% | 0 | 100.00% |
| 131,072 | LSA-128 | 64/70 (91.4%) | 82.53–96.01% | 0.000120058 | 100.00% |
| 131,072 | LSA-16 | 64/70 (91.4%) | 82.53–96.01% | 0.000119079 | 100.00% |
| 131,072 | Native QSA, replay configuration | 63/70 (90.0%) | 80.77–95.07% | 0.0488958 | 98.57% |
| 131,072 | QSA port | 65/70 (92.9%) | 84.34–96.91% | 0.00757673 | 100.00% |
Per-score intervals are Wilson binomial intervals, descriptive only; they do not account for shared record placements. Paired intervals resample the ten placements, preserving all seven task families in each cluster. Tests are exact two-sided McNemar, with Holm adjustment across the full twelve planned LSA-versus-native-QSA/dense comparisons. The family remains incomplete. First-token KLD uses the top 1,024 probabilities plus a residual bucket; it is not full-vocabulary or whole-answer KLD. Many value answers share the first token V, so this diagnostic can miss later digit errors.
| Historical 128K comparison | Difference in percentage points, paired 95% CI | Wins / losses | McNemar p | Holm p |
|---|---|---|---|---|
| LSA-128 vs native QSA replay | +1.43 [−7.14, +8.57] | 4 / 3 | 1 | 1 |
| LSA-16 vs native QSA replay | +1.43 [−7.14, +8.57] | 4 / 3 | 1 | 1 |
| LSA-128 vs dense | +1.43 [0.00, +4.29] | 1 / 0 | 1 | 1 |
| LSA-16 vs dense | +1.43 [0.00, +4.29] | 1 / 0 | 1 | 1 |
These observations show no significant advantage and do not demonstrate equivalence.
RTX 3090 Ti rerun: 192K and 255K
The new comparison uses one RTX 3090 Ti at 350 W, PCIe gen4 x16, AMD Ryzen 7 5800X3D. It is separate from the RTX 3090 screen. Every arm must reach 70 questions, with each QSA implementation's dense diagnostics on this same card. The complete five-arm comparison is pending. One supplemental native reference-mask score is available; it cannot substitute for the unresolved default cached selector.
| Tokens | Arm | Target questions | Completed validated full score | 95% CI | McNemar / Holm p |
|---|---|---|---|---|---|
| 196,608 | Dense | 70 | 61/70 (87.14%) | 77.34–93.09% | Reference |
| 196,608 | LSA-128 | 70 | 63/70 (90.00%) | 80.77–95.07% | 0.625 / 1.0 vs dense |
| 196,608 | LSA-16 | 70 | 63/70 (90.00%) | 80.77–95.07% | 0.625 / 1.0 vs dense |
| 196,608 | Native QSA | 70 | Pending, will be added | Pending | Pending |
| 196,608 | QSA port | 70 | Pending, will be added | Pending | Pending |
| 261,120 | Dense | 70 | 61/70 (87.14%) | 77.34–93.09% | Reference |
| 261,120 | LSA-128 | 70 | 61/70 (87.14%) | 77.34–93.09% | 1.0 / 1.0 vs dense |
| 261,120 | LSA-16 | 70 | 62/70 (88.57%) | 79.04–94.09% | 1.0 / 1.0 vs dense |
| 261,120 | Native QSA | 70 | Pending, will be added | Pending | Pending |
| 261,120 | QSA port | 70 | Pending, will be added | Pending | Pending |
At 196,608 tokens, the completed dense, LSA-128, and LSA-16 rows have no truncated answers. First-answer top-1,024-plus-tail KLD versus dense is 0.000330692 for LSA-128 and 0.000293658 for LSA-16; both match the dense top token on 69/70 questions (98.57%). LSA-128 minus dense is +2.86 percentage points, with a ten-placement cluster bootstrap 95% interval of [0.00, +7.14]. Exact two-sided McNemar has 3 wins and 1 loss, p=0.625; Holm p=1.0 across the 12 planned comparisons. LSA-16 has the same correct/incorrect result on every question as LSA-128, so its comparison with dense has the same paired interval and p-values. This is inconclusive and does not establish equivalence between block sizes. Score intervals in the table are descriptive Wilson intervals; these 70 synthetic questions are not an independent sample of real workloads.
The native engine's reference sparse-mask option completed a supplemental 196,608-token cell: 61/70 fully correct (87.14%; descriptive Wilson 95% CI 77.34–93.09%), with no truncated answers. First-answer top-1 agreement with its own same-card dense diagnostics was 70/70; top-1,024-plus-tail KLD was 0.0476979 (placement-cluster bootstrap 95% CI 0.0360573–0.0611926). This is a fresh test with greedy decoding and K2048 on the RTX 3090 Ti described above. The two 35-question shards used identical document, engine, model, sampler, and inference settings; the first stored its full document state in memory and the second on disk. Both passed the strict first-question live/restored raw-probability controls. Only the first question in each shard had an independent live control. The dense cells supplied probability diagnostics without generated dense recall answers, so no native QSA-versus-own-dense recall test can be calculated for this card. The full comparison remains pending. This option does not validate the published default cached selector, and its score remains outside that row above.
At 261,120 tokens, LSA-128 and dense each scored 61/70 with no truncated answers. LSA-128 corrected one dense pointer error and introduced one binding error: one win and one loss, paired difference 0.00 percentage points, placement-cluster 95% CI [−4.29, +4.29], exact McNemar p=1 and Holm p=1. First-answer KLD was 0.000612118, with top-token agreement on 69/70 questions (98.57%). This result is inconclusive; the QSA arms are still pending. LSA-16 scored 62/70 (88.57%) with no truncated answers: two wins and one loss versus dense, difference +1.43 percentage points, placement-cluster 95% CI [−4.29, +8.57], exact McNemar p=1 and Holm p=1. Its first-answer KLD was 0.000672244, with top-token agreement on 69/70 questions (98.57%). Against LSA-128, LSA-16 had two wins and one loss, a descriptive paired difference of +1.43 points [−2.86, +5.71], p=1; this extra block-size comparison was not part of the twelve planned confirmatory tests. These results do not support a block-size recommendation.
Why did a historical sparse score exceed dense?
In the earlier RTX 3090 screen, LSA-128 had five wins and no losses against dense at 196,608 tokens. Two dense failures selected the value associated with the one-character decoy key; three were arithmetic errors. At 261,120, there was one win each: a dense decoy error and an LSA arithmetic error. All seven answers were short, validly formatted, and untruncated. Recorded engine/model, document hashes, context, RoPE arguments, sampler, and seed matched within each pair; dense had no LSA environment setting or activation log.
Full attention is a reference distribution, not a guarantee of the correct answer. Changing attention changes that distribution. Sampled exact-answer correctness at temperature 1 with mb-MTP need not increase monotonically with attended-token count, even with the same seed. Different draft acceptance counts also mean that the seed alone does not establish identical random-number use. This is a plausible explanation, not a controlled separation of sampling, attention, and MTP effects. It does not prove the engine is bug-free or that sparse attention has better quality. The five-versus-zero result has exact two-sided p=0.0625 and Holm p=0.75; the one-versus-one result has p=1. Neither supports a superiority claim.
The new 196,608-token RTX 3090 Ti pair has 3 wins and 1 loss for LSA-128. Dense selected two decoy values and made one arithmetic error; LSA-128 made a different arithmetic error. All four answers were short, untruncated values under the same recorded document, sampler, seed, and model. The p=0.625 result gives no evidence that sparse recall is better.
The final recommendation between block sizes 16 and 128 remains pending. This release preserves the incomplete table and protocol qualifications; missing results will be added when validated.
Additional validation finding: the native engine's reference sparse-mask configuration failed a longer continuation check after full prompt-state restore. Live and restored top-1,024 ranks differed at 1,019 positions following the same 512-token continuation. That incomplete test contributes no recall score. The shorter controls described above do not validate this longer replay path. The 255K native result and all quality conclusions remain pending.
- Downloads last month
- -