base_model: LiquidAI/LFM3-1.7B license: other license_name: lfm-open-license-v1.0 license_link: https://huggingface.co/LiquidAI/LFM3-1.7B/blob/main/LICENSE language:
- en
- zh
- de pipeline_tag: text-generation library_name: gguf tags:
- gguf
- quantized
- lfm3
LFM3-1.7B Uncensored — GGUF (by AtomicChat)
Safety-ablated GGUF of LiquidAI/LFM3-1.7B. A projected rank-1 LoRA (direction = mean harmful − mean harmless activations, with the harmless component projected out) is baked into the weights at W*=2.0, then AD-quantized.
What changed (measured)
| metric | base | uncensored |
|---|---|---|
| refusals on harmful prompts | 61.7% | 0.0% |
| false refusals (XSTest) | 5.2% | 0.0% |
| MMLU (2000 q) | 41.0% | 41.0% (McNemar p=1.0 — no drop) |
| needle (long context) | 3/3 | 3/3 |
| ΔKLD vs base (neutral) | — | +0.019 |
A random edit of the same size dropped MMLU 1.3pp and had a larger ΔKLD (0.028), so the targeted direction is both more effective and less damaging. Thinking-loop and empty-output checks pass at W*=2.0.
Rungs
| Rung | GB on disk | bpw |
|---|---|---|
| AD-Q6_K | 1.6 | 7.31 |
| AD-Q4_K_M | 1.1 | 5.20 |
Run (llama.cpp)
Needs a build with LFM3 support (merged in llama.cpp #31252; produced at llama.cpp 19cce69).
llama-cli -hf AtomicChat/LFM3-1.7B-Uncensored-GGUF:AD-Q4_K_M -p "Hello"
Base model & license
Base model: LiquidAI/LFM3-1.7B. License: LFM Open License v1.0 (see LICENSE). This removes safety refusals; use responsibly and in line with the base license.
Quants by AtomicChat.