Quantized builds of alphalabz/LYNX-ALPHA, an offensive-security model for red teaming, penetration testing, OSINT, malware analysis, and exploit development. Measured on the same harness as the fp16 release, identical prompts, greedy decoding: Q4KM gives up 1 to 2 points, which is the ordinary cost of 4-bit weight quantization and keeps the model inside its base model’s confidence interval everywhere. It also drops one additional refusal (5/12 → 4/12). The f16 build these were quantized from passed a structural gate before conversion: blockcount=32`, 427 tensors, no phantom block from the multi-token-prediction head. Digit probes pass on the GGUF itself, not just on fp16. The model was trained with thinking disabled. If your runtime supports it: Temperature 0.7, top_p 0.9, repeat penalty 1.1 works well for security Q&A. For exact-answer tasks use temperature 0. Same as the fp16 release: single training run, capped benchmark subsets, weak refusal measurement, hallucination on specific CVE numbers and version strings, English only, no tool-use training. Full detail and confidence intervals are in the model card and the paper. Licensed under Apache 2.0, inherited from the base…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: alphalabz
Теги: gguf, llama.cpp, ollama, security, offensive-security, red-teaming, uncensored, qwen3.5
Лайков: 3 | Загрузок: 1,264
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
