folfkspdvvk/LFM2.5-1.2B-Thinking-Gemini-Pro-1000-Heretic-Uncensored-DISTILL
folfkspdvvk/LFM2.5-1.2B-Thinking-Gemini-Pro-1000-Heretic-Uncensored-DISTILL is a 1.2 billion parameter language model, fine-tuned from LFM2.5-1.2B, specifically designed for deep and detailed reasoning. This model features a 128k context window and is uncensored, having been "Heretic'ed" before tuning to ensure direct and unconstrained output. It excels at generating compact yet highly detailed reasoning and is optimized for use with specific quantization levels and sampling settings to prevent looping.
Loading preview...
Model Overview
folfkspdvvk/LFM2.5-1.2B-Thinking-Gemini-Pro-1000-Heretic-Uncensored-DISTILL is a 1.2 billion parameter model, fine-tuned from the LFM2.5-1.2B base, with its reasoning capabilities entirely replaced using distillation from advanced reasoning datasets. This model is notable for its "Heretic" uncensored nature, meaning it was de-censored prior to fine-tuning to ensure direct and unconstrained content generation without refusals.
Key Capabilities & Features
- Enhanced Reasoning: The model's core strength lies in its compact yet highly detailed reasoning, which influences general operation, output generation, and benchmarks.
- Extended Context: Features a substantial 128k context window, allowing for processing longer inputs and maintaining coherence over extended interactions.
- Temperature Stability: Reasoning remains stable across a wide temperature range of 0.1 to 2.5.
- Uncensored Output: Designed to generate content without refusal, though specific directives may be needed for highly graphic or explicit content to meet desired intensity levels.
Optimal Usage & Settings
For best performance, the model recommends using q5, q6, q8, or 16-bit precision, or Imatrix IQ3_M minimum. A repetition penalty of 1.05 to 1.1 is suggested. To prevent looping during thinking, especially with lower quants, users should lower the temperature to 0.3-0.7 or increase the repetition penalty. The model also benefits from a "Smoothing_factor" of 1.5 in interfaces like KoboldCpp, oobabooga, or Silly Tavern for smoother operation in chat and roleplay scenarios.