!image This repository contains a surgically de-censored version of Syntropy AI’s Soren-1-Small model, optimized via weight abliteration techniques. By applying the heretic framework across an extensive 200-trial search space, we successfully isolated and neutralized the primary refusal vectors embedded within the attn.oproj and mlp.downproj layers. — Optimal Balance: Selected Trial 156 out of 200 iterations for the perfect trade-off between freedom and reasoning capabilities. — Refusal Rate: Dropped down to 5/100 (from the original near-total refusal on safety benchmarks). — KL Divergence: 0.0380 — Demonstrates that general language capabilities are preserved relative to the original model. However, safety-aligned weights in attn.oproj and mlp.downproj have been surgically removed; this is intentional modification, not unintended degradation. Benchmark evaluations could not be completed due to the model’s weight constraints and available compute resources. Standardized benchmarks (GSM8K, HellaSwag, MMLU) will be added soon once proper evaluation infrastructure is set up. Unlike aggressive fine-tuning which often degrades the model’s core logic or shifts its grammar distributions,…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: K0D3IN
Теги: qwen3_5_text, soren, soren-1-small, heretic, abliteration, apache2, conversational, en
Лайков: 3 | Загрузок: 13
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
