Datasets:
Evidence required for a SOTA claim
Eve's Embrace of Extreme Quantization v1.0 is a complete research prototype release. The name does not certify SOTA. A first-block OPT-125M diagnostic, calibration objective improvement, a sharp mathematical bound, and a very short constructed synthetic codeword are different evidence categories.
What this release can establish
- Correct strict finite formats within their documented contract and checked cases.
- Exact proofs under the hypotheses stated in the mathematical appendix.
- Complete byte counts for the actual six-projection bundle and corrections.
- Native FP32 quality for the frozen released test window and installed states.
- Protection of bare candidates in the supplied finite calibration catalog.
- Reproducible source, runtime, model and token provenance.
What a leadership claim would additionally require
| Dimension | Required evidence |
|---|---|
| Model scope | All relevant projection weights, embeddings and auxiliary state; contemporary dense and mixture-of-experts model families and multiple sizes |
| Quality | Standard full-corpus perplexity, several language and reasoning tasks, generation behavior, and explicit float reference gaps |
| Comparators | Current NanoQuant, AF1, QTEA, strong two-bit methods and applicable deployment baselines, using their documented recipes |
| Fairness | Comparable calibration data, optimization access and encoder compute; no uncharged adapters, codebooks, routing, exceptions or reconstruction state |
| Rate | Actual complete deployed bytes, including all scales, supports, finite factors, directories, alignment and overlays |
| Execution | Real packed full-model inference; measured latency, throughput, peak memory, batch/context dependence and target hardware |
| Generality | Multiple calibration seeds, corpora, model families and workloads; transparent failures and uncertainty |
| Independence | Reproduction by someone other than the implementing author, source/evidence integrity and current prior-art audit |
No numerical completion percentage estimates the probability that this method is best in the world. No calibration average supplies a uniform autoregressive-tree certificate. Future studies must reserve new evaluation data before fitting or method selection. The released final window is available for replication, but cannot become fresh again.
The primary and extreme profiles are selected by an immutable actual-byte-constrained calibration rule. The selection may choose a bare numerical control or a factor reserve. A source-specific hybrid result is disclosed separately when the global winner differs. The release does not force the fork to win, replace the default after seeing fresh scores, or interpret a local GPTQ-style control as an industry-wide benchmark.