Lucet-Coder-Omega is post-trained from XHToken/Spark-X2.6-4B (1M native context window, 54.3% SWE Pro, 71.8% BFCL-V4) using four advanced mathematical and systems-level algorithms: 1. 1M-Context RoPE-Preserving Split-Stack Adaptation (r=16): Keeps qproj and kproj 100% frozen across all layers (preserving the 1,048,576-token RoPE phase calibration) and adapts only Value (vproj), Output (outproj), and MLP (gateproj, upproj, downproj, gproj) matrices in the upper 50% of the network. 2. Failure-Cluster Active Data Mining & AST Counterfactual Bug Immunization: Diagnoses failure modes (FAILTRUNCATEDTHINK, FAILASSERTIONBOUNDARY, FAILMISSINGIMPORT, FAILEPISTEMICHALLUCINATION) across algorithmic clusters and dynamically pops unseen neighbor tasks + surgical AST counterfactual bugs from a disjoint reserve pool—achieving +14.0% on Held-Out MBPP pass@1 and +20.1% faster token efficiency with zero training data repetition. 3. QR-SVD Gavish-Donoho Spectral Denoising: Strips stochastic gradient noise singular values from low-rank adapters in $O(r^2(d{text{out}} + d{text{in}}))$ before merging into base weights. 4. Q4K 32-Block Crest-Factor Equalization & KL-Bounded Uncensored-GSQ-v2: Equalizes…
Модальности:
Генерация текста
Области применения:
Диалог / чат Генерация кода
Задача: Генерация текста
Автор: mohommed
Теги: spark2_5, lucet, spark-x2.6, qr-svd, ast-counterfactual, failure-cluster-mining, uncensored, gsq
Лайков: 3 | Загрузок: 489
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
