license_name: lfm1.0
license_link: LICENSE
license: other
base_model:
- LiquidAI/LFM3-1.7B
tags: - Abliterated
- GGUF
LFM3-1.7B-Abliterated-GGUF
GGUF builds of an abliterated derivative of LiquidAI/LFM3-1.7B (1.73B parameters, 24-layer tri-hybrid: gated short-conv + sliding/full attention).
Benchmarks
All numbers below were measured on a user-provided Japanese evaluation set (84 harmful/NSFW prompts and 80 knowledge questions). Refusal labels are string-match heuristics, not human labels.
Refusal (84 prompts, thinking ON)
| Model | Refusals |
|---|---|
| LFM3-1.7B (base) | 40 / 84 |
| This model | 0 / 84 |
Knowledge (80 questions, JA/EN, thinking OFF)
| Model | Correct |
|---|---|
| LFM3-1.7B (base) | 49 / 80 |
| This model | 48 / 80 |
Quantization (llama.cpp llama-perplexity, 40 chunks)
| File | PPL | vs. BF16 | Refusals (10 prompts, greedy) |
|---|---|---|---|
| BF16 | 6.8849 | — | 0 / 10 |
| Q8_0 | 6.8915 | +0.10% | 0 / 10 |
| Q6_K | 6.8820 | −0.04% | 0 / 10 |
| Q4_K_M | 7.0406 | +2.3% | 0 / 10 |
Files
| File | Size | Notes |
|---|---|---|
| LFM3-1.7B-Abliterated-BF16.gguf | 3.47 GB | Original weights |
| LFM3-1.7B-Abliterated-Q8_0.gguf | 1.85 GB | Recommended (practically lossless) |
| LFM3-1.7B-Abliterated-Q6_K.gguf | 1.49 GB | Smaller, PPL on par with BF16 |
| LFM3-1.7B-Abliterated-Q4_K_M.gguf | 1.16 GB | Smallest, +2.3% PPL |
Run (llama.cpp)
Requires llama.cpp b12040 or newer. Older builds fail with unknown model architecture: 'lfm3'.
Notes:
- Thinking is ON by default (the chat template starts the assistant turn with thinking). Disable it with
"chat_template_kwargs": {"enable_thinking": false}. - Use
max_tokens >= 4096. Thinking is long (median ~1.7k tokens on the eval set); with 2048, the answer is often cut off mid-thought. - Sampling used for the benchmarks: temperature 0.1, top_k 50, top_p 1.0, repetition_penalty 1.1 (as recommended in the upstream model card).
Changes
- Weight edit only. No training, fine-tuning, or distillation.
- Removed a refusal direction from 40 residual-writer matrices in layers 4–23:
self_attn.o_proj,conv.out_proj, andfeed_forward.w2. - Direction: per-layer mean difference of hidden states between harmful and harmless prompts (416 pairs), taken at the last prompt token. The projection is computed in FP32, and the weights are saved in BF16.
- Orthogonalization is sandwich-norm aware: this model adds each writer's output to the residual after
post_operator_norm/post_ffn_norm. The projection therefore targetsg ⊙ u(g = the norm's learned scale) instead ofu, so the removed component does not return after the norm. - Untouched:
embed_tokens, all norms,q/k/v_proj,conv.in_proj,conv.conv,attn.gate_proj,w1/w3,lm_head(tied). - Relative weight change: mean 1.87e-04, max 3.53e-04; max |ΔW| = 2.97e-01.
License
Released under the LFM Open License v1.0 (copy included as LICENSE). This is a derivative work of LiquidAI/LFM3-1.7B; the files were modified, as stated in NOTICE-MODIFICATIONS.md.
- Commercial use is limited to entities with annual revenue below 10 million USD.
- Liquid AI does not endorse or support this model.
- You are responsible for how you use it. Do not use it to generate sexual content involving minors, to assist with illegal acts, or to target real people.