AtomicChat/LFM3-1.7B-Uncensored-GGUF
AI model · https://huggingface.co/AtomicChat/LFM3-1.7B-Uncensored-GGUF
First seen by Oriel on .
AI summary
This Hugging Face page presents an AD-quantized, safety-ablated GGUF of LiquidAI/LFM3-1.7B that removes measured refusals without reducing MMLU or needle performance.
- A projected rank-1 LoRA targeting harmful-versus-harmless activation directions was baked into the weights at W*=2.0 before AD quantization.
- Harmful-prompt refusals fell from 61.7% to 0.0%, while false refusals on XSTest fell from 5.2% to 0.0%.
- MMLU remained 41.0% across 2,000 questions, needle performance remained 3/3, and neutral ΔKLD versus base was +0.019.
- Available rungs are AD-Q6_K at 1.6 GB and 7.31 bpw, and AD-Q4_K_M at 1.1 GB and 5.20 bpw.
- Running it requires llama.cpp with LFM3 support, merged in llama.cpp #31252 and produced at llama.cpp 19cce69.
#gguf #llama.cpp #model quantization #language models
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Likes
Total likes reported by the source.
- Hugging Face Models: 273, latest observation ; +161 over observed period since
Downloads (30d)
Rolling 30-day downloads. A change in this value is not new downloads during the observed period.
- Hugging Face Models: 103,039, latest observation ; +100,722 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Models | AtomicChat/LFM3-1.7B-Uncensored-GGUF | https://huggingface.co/AtomicChat/LFM3-1.7B-Uncensored-GGUF Read discussion |