Cadence 0.8B — qualified checkpoint2357
This release contains the learned FP32 rank8 adapter and pointer head for a758,329,152-parameter decision model, plus portable CUDA inference code. It produces distributions for choice, boolean/noul, and ordinal score questions. It does not generate text. One question is accepted per wrapper request. It is a custom two-stage adapter/head format: use the supplied loader, not a generic PEFT-only or AutoModel loader.
The audited 17-file release is published at foolish-bandit/cadence-0.8b, with release commit de5009eeca68c08110e12bcd44c08016c5fe871d.
Setup and use
The qualified CUDA profile was tested with Python3.12+, Torch2.8.0+cu128, transformers5.19.0, peft0.21.2, safetensors0.8.0 and numpy2.5.4. Install a compatible CUDA-enabled Torch2.8.0 wheel, then the pinned core dependencies:
python -m pip install torch==2.8.0 --index-url https://download.pytorch.org/whl/cu128
python -m pip install -r requirements.txt
python cadence_runtime.py --request example_request.json
For repeated requests, load once:
from cadence_runtime import Cadence
model = Cadence.from_pretrained('.')
result = model.predict({
'state': 'A delivery arrived two days after its promised date.',
'questions': {'decision': {
'type': 'choice', 'instructions': 'Classify delivery timing.',
'criteria': {'on_time': 'No later than promised.', 'late': 'After the promised date.'}
}}
})
print(result['probabilities'])
After publication, first download this package at its exact release commit with huggingface_hub.snapshot_download, add that downloaded directory to Python's module path, and import its cadence_runtime. Cadence.from_pretrained also accepts a Hub repository ID when an exact revision is supplied. Review the code before running it. CPU and MPS are not validated and the loader rejects them. This package is portable across suitable CUDA installations with these dependencies; it is not a standalone CPU model or a copy of the full base weights.
The loader fetches public upstream dependencies at exact immutable revisions, or reuses their normal Hub caches. Fresh use needs the Qwen base weights separately, approximately1.75GB, and about45MB of released Kev adapter/head dependencies. This release itself is about24MB. Offline use requires those pinned snapshots already cached; set HF_HUB_OFFLINE=1 and use local_files_only=True. Do not substitute newer base or adapter revisions.
Exact reconstruction
- Load Qwen/Qwen3.5-0.8B-Base atdc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68.
- Apply and merge jaredpalmer/kev-0.8b at9a45d25eb2ab761841196625383fa1dff0e56c1e into BF16 frozen base weights using unmodified pinned Kev code.
- Construct the rank8/alpha16 adapter with the exact target modules and copy all376 qualified FP32 adapter/head tensors. Require exact shape, precision, inventory and post-copy equality.
- Use eval, inference_mode, SDPA math, TF32 off, no CUDA graphs, BF16 forward autocast and T1. The loader creates no optimizer and disables gradients.
The files are adapter.safetensors, head.safetensors, cadence_config.json, cadence_runtime.py, cadence_kev inference modules, qualification_metrics.json, example_request.json, requirements.txt, LICENSE and NOTICE. No optimizer states, raw corpus rows, local project paths, credentials or private notes are included. Weights and upstream files are SHA-checked.
Requests preserve semantic choice IDs and descriptions, including null descriptions for native full150-class CLINC. Score criteria are an ascending list; keys are numeric level IDs. Noul returns P(true). Returned vectors use native key order; align by semantic ID rather than positional assumptions. No extra temperature is required. Overflow is rejected at8192 state tokens,8192 branch tokens,16384 packed tokens. Training used1024 and final evaluated inputs reached686tokens; larger-context performance is not established.
Qualification
The exact fixed suite has6933 rows in4196 known source groups: WiC754, SNIPS521, CLINC503, MNLI3039, STS1114 and SST1002. All five families were seen during training. Candidate2357 was preselected and frozen before final inference; no final training, calibration fitting, prompt selection or reselection occurred.
| Model | Full-suite equal-family J |
|---|---|
| candidate2357 | 0.8255106815289301 |
| full857 | 0.792424232891669 |
| originalKev | 0.7513046995517388 |
Gain over857: 0.03308644863726118, paired two-sided95% CI [0.02160043186609384, 0.044732048822041404]. Gain over shippedKev: 0.0742059819771913, CI [0.058568039826161544, 0.08909689077721061]. Both prespecified qualifications pass:857 gain≥0.01 and all safeguards;Kev gain≥0.03; strict positive lower bounds and zero undefined required draws. These are marginal intervals, not simultaneous95% coverage.
| Native task metric | Candidate2357 | Prior857 | Shipped Kev | Default439 |
|---|---|---|---|---|
| wic accuracy | 0.6684350132625995 | 0.6737400530503979 | 0.5305039787798409 | 0.48010610079575594 |
| wic macro_f1 | 0.664848372781065 | 0.6613896089671036 | 0.5089411868597669 | 0.38538573508005824 |
| snips accuracy | 0.9692898272552783 | 0.944337811900192 | 0.8541266794625719 | 0.9500959692898272 |
| clinc150 accuracy | 0.8489065606361829 | 0.6500994035785288 | 0.6481113320079522 | unsupported |
| multinli accuracy | 0.8265876933201711 | 0.7976307996051333 | 0.805528134254689 | 0.5294504771306351 |
| stsb pearson | 0.9054374483250518 | 0.8695493784096272 | 0.8255993493874889 | 0.5112309858113304 |
| stsb spearman | 0.8912655653286139 | 0.8533693259079963 | 0.8215355544760224 | 0.5687803323509473 |
| sst5 accuracy | 0.5568862275449101 | 0.562874251497006 | 0.5668662674650699 | 0.5089820359281437 |
| sst5 normalized_expected_MAE | 0.1347935916453442 | 0.13036762593707377 | 0.13908867942851175 | 0.14209409396274683 |
| sst5 normalized_RPS | 0.08856245334676313 | 0.10704920935790789 | 0.09032325320726324 | 0.09615772429238649 |
Normalized expected MAE and RPS are lower-is-better. All five family point scores improve versus857, but SST accuracy is slightly below both857 andKev, and the sentiment family is slightly belowKev. All frozen857 regression safeguards pass. Default439 has6430 compatible rows and503 explicit CLINC incompatibilities; no full439 J, intent score or renormalized subset J is reported. Compatible task intervals in qualification_metrics.json are descriptive; SNIPS accuracy includes zero and expected SST MAE crosses zero.
J is the mean of five equally weighted family scores: WiC macroF1; mean SNIPS/CLINC accuracy; MNLI accuracy; (1+mean STS Pearson/Spearman)/2; mean SST accuracy,1-normalized expected MAE,1-normalized RPS. Native row/class populations are intact. Paired bootstrap uses10000 draws, seed20261012, stdlibMT19937 replacement sampling of lexically sorted known source groups independently within each dataset, applying identical group multiplicities to compared models. Nearest-rank2.5%/97.5% endpoints; undefined required draws reject intervals. Original857 guards additionally require≥3 positive families, family/native score regressions≤0.03 and ordinal loss increases≤0.02, complete finite native coverage and preserved recovery/retention diagnostics. Independent verification reconstructed original gold and reproduced268 statistical scalars within5.55e-16, with exactly matching gate booleans and zero selection epsilon.
This supports only the defined training-seen-family group holdout. It does not prove unseen-task, general-purpose, multilingual, long-context, best-in-class or pretraining-clean superiority. Unknown deeper source ancestry includes SST review/tree, SNIPS/CLINC prompt and MNLI document/hypothesis relationships; public-base pretraining exposure is unknown. The32causal/192numeric TRAIN retention cases are diagnostics, not human-generalization evidence. The additional comparison below uses this already examined suite and is descriptive.
Additional comparison on the already examined suite
The qualified checkpoint2357 remained frozen. This descriptive comparison reused the same 6,933 rows and original Cadence/Kev forecasts; it is not a fresh blind qualification or a model selection round. Accuracy and 10,000 whole-source-group paired bootstrap draws have passed independent CPU verification. The original qualification and its gates are unchanged.
| Metric | Cadence2357 | Shipped Kev | GLiClass instruct | Laya | Qwen3.5-0.8B |
|---|---|---|---|---|---|
| WiC accuracy | 0.668435 | 0.530504 | 0.446950 | 0.551724 | 0.553050 |
| WiC macro F1 | 0.664848 | 0.508941 | 0.308891 | 0.541610 | 0.356106 |
| SNIPS accuracy | 0.969290 | 0.854127 | 0.882917 | 0.802303 | 0.664107 |
| CLINC150 accuracy | 0.848907 | 0.648111 | 0.532803 | NA | 0.385686 |
| MNLI accuracy | 0.826588 | 0.805528 | 0.288911 | 0.903587 | 0.343534 |
| STS mean Pearson/Spearman | 0.898352 | 0.823567 | 0.255312 | 0.701911 | NA |
| Native SST argmax accuracy | 0.556886 | 0.566866 | 0.435130 | 0.370259 | NA |
| SST scalar nearest-anchor accuracy | 0.562874 | 0.565868 | 0.446108 | 0.372255 | 0.397206 |
All accuracies are fractions. Qwen has no native class-probability vector or native SST argmax; its scalar nearest-anchor accuracy is reported separately, with lower-anchor tie breaking. Seven Qwen CLINC parse failures remain in the full 503-row denominator. Its STS correlations and scalar MAE are NA because one resource-interrupted row lacks a prediction; no favorable valid-only subset is substituted. There were 6,925 valid Qwen outcomes, seven parse failures and one indeterminate interruption, with no native replay.
| Aggregate | Value |
|---|---|
| Cadence J_native | 0.825511 |
| Kev J_native | 0.751305 |
| GLiClass J_adapted | 0.528888 |
Laya has 503 preflight-unsupported CLINC rows and no full, intent or renormalized subset J. Qwen has no J or probability-based RPS. GLiClass uses an explicit typed-question adaptation; J_adapted is distinguished from native J. A saved-row and pinned SDK audit found no label-order or argmax mapping defect in its low MNLI result: 2,514 neutral predictions, 525 entailment and zero contradiction predictions. This result applies to this fixed adapted setup.
| Paired difference, Cadence minus comparator | Point | Marginal two-sided 95% CI |
|---|---|---|
| J_native: Cadence minus Kev | 0.074206 | [0.058568, 0.089097] |
| J_native minus GLiClass J_adapted | 0.296622 | [0.283363, 0.310423] |
| MNLI accuracy: Cadence minus Laya | -0.076999 | [-0.091059, -0.063196] |
| WiC accuracy: Cadence minus Qwen | 0.115385 | [0.024142, 0.206785] |
| SNIPS accuracy: Cadence minus Qwen | 0.305182 | [0.251984, 0.364486] |
| CLINC accuracy: Cadence minus Qwen | 0.463221 | [0.417495, 0.508946] |
| MNLI accuracy: Cadence minus Qwen | 0.483054 | [0.469622, 0.496246] |
| SST nearest-anchor accuracy: Cadence minus Qwen | 0.165669 | [0.128871, 0.202393] |
Intervals use the frozen 10,000 paired group draws, seed20261012 and unchanged class/row denominators. They are descriptive marginal intervals, not simultaneous 95% coverage. The independent verifier recomputed every point estimate and denominator, regenerated the complete draw stream, and independently recomputed 14 WiC/MNLI accuracy intervals; remaining interval arithmetic was source-reviewed and pinned. Laya leads MNLI, and Kev is slightly ahead on native SST accuracy; no general-purpose or best-in-class superiority claim follows.
Comparator profiles and latency
GLiClass: knowledgator/gliclass-instruct-large-v1.0 at825e5478c1bf4bffbf297690517097ccbdb2e006, official GLiClass0.1.20, FP32, context1024, T1. Laya: convaiinnovations/laya at7b928d828b7b0e022f929d9bd2e44165aa270148, official SDK0.4.6, FP32 parameters with BF16 forward autocast and frozen native task temperatures. Qwen: Qwen/Qwen3.5-0.8B at2fc06364715b967f1860aea9cf38778875588b17, BF16, official nonthinking template, greedy one-beam decoding capped at32 new tokens, strict whole-answer JSON parsing. These profiles differ in architecture, precision, calibration and interface.
Fixed synthetic latency panel
Batch1, NVIDIA GeForce RTX3050OEM. Each cell shows median / p95 milliseconds across all20 timed native encode/infer/CPU-transfer calls; three fixed warmup attempts precede each case. Loading, per-row guards, journaling and audits are excluded. Models retain their declared precision/calibration/kernel profiles; this is not a universal serving-speed or architecture-only comparison.
| Fixed case | Cadence2357 | Kev | GLiClass | Laya | Qwen |
|---|---|---|---|---|---|
| synthetic:choice:2 | 81.195 / 113.688 | 46.421 / 57.607 | 41.168 / 57.278 | 27.124 / 38.884 | 211.270 / 258.517 |
| synthetic:choice:7 | 62.765 / 139.731 | 57.742 / 61.029 | 37.931 / 53.220 | 27.496 / 33.301 | 298.587 / 327.912 |
| synthetic:noul:2 | 59.561 / 156.696 | 42.801 / 66.961 | 37.578 / 45.837 | 26.995 / 27.796 | 196.153 / 281.275 |
| synthetic:score:6 | 59.304 / 133.040 | 45.335 / 51.588 | 36.232 / 56.132 | 29.737 / 61.287 | 191.660 / 214.078 |
| synthetic:choice:150 | 356.288 / 363.687 | 447.885 / 462.260 | 223.423 / 229.509 | NA (unsupported) | 877.776 / 924.731 |
| Profile | Cold load seconds | Observed peak allocated / reserved GiB |
|---|---|---|
| Cadence2357 | 249.446 | 1.591 / 1.791 |
| Kev | 13.117 | 2.997 / 3.197 |
| GLiClass | 12.148 | 1.850 / 1.889 |
| Laya | 8.614 | 1.591 / 1.693 |
| Qwen | 6.064 | 2.135 / 2.342 |
Cadence first loaded in 9.260s before the observer failure; the table reports its resumed cold load. The first warmup produced a native prediction through one source-confirmed direct forward, which the original module pre-hook missed. That immutable attempt was not replayed. Its timing and memory telemetry are unavailable; the recovered first fixture spans two model constructions and has two warmups in the resumed process. All20 timed repetitions remain intact. Reported memory is the maximum observed telemetry, not an exact bound for the unobserved first warmup.
Cold-load times include pinned-artifact validation and model reconstruction. Cadence's resumed 249.446 s load involved substantial disk reads, while its original segment loaded in 9.260 s. Background CPU/disk activity and cache state were not controlled; these are observed profile timings, not a stable startup or service SLA.
The repaired baseline observer passes arguments and outputs through unchanged and counts at method entry, including direct forward calls. Other actors keep the original module pre-hook. GPU processes ran sequentially under the comparison supervisor; unrelated training was left alone. Qwen output statuses and generated-token counts are included in the accompanying metrics rather than assuming every timed generation was valid.
Precision and timing context
Candidate and857 use a BF16 frozen merged base with FP32 adapter/head, T1. ShippedKev was measured in FP32 with nativeT2.3510958125672174 already applied; default439 uses FP16,512context, nativeTchoice1.1731808243906614/noul1.0004287566844705/score1.0095375299959983. These profiles are not harmonized; comparison does not isolate architecture from precision or calibration.
Original6933 candidate native encode/infer/CPU-transfer timers on RTX3050OEM8GiB: median0.043280172045342624s; p950.22219638602109626s; mean0.05760373803917907s; sum399.3667158256285s. Variable inputs/options, batch1, reference PyTorch kernels; loading, journal writes and audits excluded. These are not an SLA or a matched-hardware serving benchmark.
Attribution, licensing and corpus provenance
Apache-2.0 for this package's new contributions; the included upstreamLICENSE retains Copyright2026JaredPalmer. Vendored Kev api/model/checkpoint files are unchanged at GitHubcommit45923b7a3460b6d36358e2e143455902c1eb856b. NOTICE identifies upstream authors and exact dependencies; no endorsement is implied. The pinned Kev and Qwen model cards declare Apache-2.0; Qwen's pinned LICENSE retains AlibabaCloud attribution. Public dependencies are downloaded separately rather than bundled.
- https://github.com/jaredpalmer/kev/tree/45923b7a3460b6d36358e2e143455902c1eb856b
- https://huggingface.co/jaredpalmer/kev-0.8b/blob/9a45d25eb2ab761841196625383fa1dff0e56c1e/README.md
- https://huggingface.co/Qwen/Qwen3.5-0.8B-Base/blob/dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68/LICENSE
Benchmark/training corpus provenance is separate from model/code licensing. This package contains no corpus rows and does not relicense datasets:
- WiC: https://pilehvar.github.io/wic/
- SNIPS: https://github.com/snipsco/nlu-benchmark
- CLINC150: https://github.com/clinc/oos-eval (CC-BY3.0 source notice)
- MultiNLI: https://cims.nyu.edu/~sbowman/multinli/ (mixed original source terms)
- STS-B: https://huggingface.co/datasets/nyu-mll/glue
- SST5: https://nlp.stanford.edu/sentiment/
Consult each original source's terms for data use or redistribution. No blanket Apache corpus license or unrestricted corpus redistribution is claimed.