AtomicChat publishes safety-ablated LFM3-1.7B GGUF
Original title: AtomicChat/LFM3-1.7B-Uncensored-GGUF — new model trending #30 on Hugging Face
AtomicChat reports harmful-prompt refusals fell from 61.7% to 0% with MMLU unchanged at 41.0%.
AtomicChat released LFM3-1.7B-Uncensored-GGUF, a safety-ablated quant of LiquidAI/LFM3-1.7B trending on Hugging Face. A projected rank-1 LoRA scaled at W*=2.0 is baked into the weights, then AD-quantized. The card says harmful-prompt refusals drop from 61.7% to 0.0% while MMLU on 2,000 questions stays at 41.0%. Files are AD-Q4_K_M (1.1 GB) and AD-Q6_K (1.6 GB); inference needs a llama.cpp build with LFM3 support.
Unconfirmed: Refusal and benchmark figures are AtomicChat's own measurements and are not independently verified.
- XSTest false refusals fell from 5.2% to 0.0%, per the model card.
- AD-Q4_K_M is 1.1 GB at 5.20 bpw; AD-Q6_K is 1.6 GB at 7.31 bpw.
- The card cites McNemar p=1.0 and no MMLU drop versus the base.
- Inference needs llama.cpp with LFM3 support from merge #31252.
- License is LFM Open License v1.0; quants are by AtomicChat.
- Products
- llama.cpp
- Organizations
- AtomicChatLiquidAI
From the original article
1 min read at huggingface.cotext-generation · 38,382 downloads · 147 likes · license other
By AtomicChat · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.