richardyoung/ lfm2.5-1.2b-instruct-heretic:Q4_K_M

7 yesterday

Uncensored LFM2.5-1.2B-Instruct via Heretic: 3/100 refusals (from 98), KL 0.059. Tiny and fast; runs on almost anything. 

ollama run richardyoung/lfm2.5-1.2b-instruct-heretic:Q4_K_M

Details

yesterday

9ed3ca55c32b · 731MB

model
lfm2
·
1.17B
·
Q4_K_M

Readme

LFM2.5-1.2B-Instruct-Heretic

Uncensored LFM2.5-1.2B-Instruct via Heretic: 3⁄100 refusals (from 98), KL 0.059. Tiny and fast; runs on almost anything.

🚀 Overview

This is an abliterated build of LiquidAI/LFM2.5-1.2B-Instruct, Liquid AI’s 1.2B edge model with a hybrid convolution + attention architecture, built for fast on-device inference. Refusal behavior was reduced using the Heretic library with KL-targeted parameters that preserve the model’s coherence.

📊 Abliteration Results

Metric Before After
Refusals 98⁄100 3⁄100
Reduction – 97%
KL Divergence – 0.059

The very low KL divergence (0.059, far below the 0.5 “damage” threshold) means the model retains essentially all of its original capabilities and coherence.

🎯 Key Features

  • Reduced censorship: 97% fewer refusals on typical “unsafe” prompts
  • Near-zero quality loss: KL 0.059
  • Tiny footprint: runs comfortably on CPU or any GPU
  • 128K-token context
  • Fast hybrid conv + attention architecture
  • Reproducible: full reproduction data on Hugging Face

🏷️ Available Versions

Tag Size BPW Notes
latest / Q4_K_M 0.7 GB 4.85 Recommended
Q5_K_M 0.8 GB 5.68 Higher quality
Q6_K 1.0 GB 6.56 Very high quality
Q8_0 1.2 GB 8.5 Near-lossless

💻 Quick Start

ollama run richardyoung/lfm2.5-1.2b-instruct-heretic           # recommended (Q4_K_M)
ollama run richardyoung/lfm2.5-1.2b-instruct-heretic:Q8_0      # near-lossless

🛠️ Use Cases

  • On-device and edge assistants without stock refusals
  • Research and red-teaming on small models
  • Creative writing

📋 System Requirements

VRAM Recommended tier
4 GB+ Q4_K_M (0.7 GB file)
4 GB+ Q8_0 (1.2 GB file)

🔧 Technical Details

  • Base Model: LiquidAI/LFM2.5-1.2B-Instruct
  • Parameters: 1.2B (lfm2 architecture)
  • Context Length: 125K tokens
  • Quantization: GGUF via llama.cpp (text generation)
  • Abliteration: Heretic v2.0.0.dev0 by p-e-w (Trial 78: 98→3 refusals @ KL 0.059)
  • License: LFM Open License v1.0 (inherited from the base model)

⚠️ Disclaimer

This model has reduced safety guardrails. The removal of refusal behavior means it will engage with a wider range of prompts. Use responsibly and in accordance with applicable laws and regulations.

🙏 Acknowledgments

  • Base Model: Liquid AI
  • Abliteration: Heretic by p-e-w
  • Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro