license_name: lfm1.0
license_link: LICENSE
license: other
base_model:
- LiquidAI/LFM3-VL-2B
tags: - Abliterated
- image-text-to-text
LFM3-VL-2B-Abliterated-GGUF
GGUF builds of an abliterated derivative of LiquidAI/LFM3-VL-2B
Benchmarks
All numbers below were measured on a user-provided Japanese evaluation set (84 harmful/NSFW prompts and 8 benign prompts).
Refusal labels come from a separate 3B decision model (d1-3B), not from string matching.
Sampling: temperature 0.1, top_k 50, top_p 1.0, repetition_penalty 1.1, thinking ON, max_new 4096.
Refusal (84 prompts, thinking ON)
| Model | Refusals |
|---|---|
| LFM3-VL-2B (base) | 58 / 84 |
| This model (BF16) | 1 / 84 |
Quantization (llama.cpp llama-perplexity, shared JA/EN corpus, 512 ctx)
| File | PPL | vs. BF16 | Refusals (84 prompts) |
|---|---|---|---|
| BF16 | 38.78 | — | 1 / 84 |
| Q8_0 | 38.74 | −0.1% | 0 / 84 |
| Q6_K | 38.50 | −0.7% | 0 / 84 |
| Q4_K_M | 42.17 | +8.7% | 2 / 84 |
Files
| File | Size | Notes |
|---|---|---|
| LFM3-2B-VL-Abliterated-BF16.gguf | 3.47 GB | Original weights |
| LFM3-2B-VL-Abliterated-Q8_0.gguf | 1.85 GB | Recommended (practically lossless) |
| LFM3-2B-VL-Abliterated-Q6_K.gguf | 1.43 GB | Smaller, PPL slightly below BF16 |
| LFM3-2B-VL-Abliterated-Q4_K_M.gguf | 1.10 GB | Smallest, +8.7% PPL and 2 refusals |
| mmproj-LFM3-2B-VL-Abliterated-BF16.gguf | 379 MB | Vision projector — required for image input (pairs with every quant above) |
Run (llama.cpp)
Requires llama.cpp b12040 or newer. Older builds fail with unknown model architecture: 'lfm3'.
llama-server -m LFM3-2B-VL-Abliterated-Q8_0.gguf \
--mmproj mmproj-LFM3-2B-VL-Abliterated-BF16.gguf \
--jinja --reasoning-format none -ngl 99 -c 8192
Notes:
- Without
--mmprojthe model runs text-only; image inputs are rejected. - 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 ~3.4k 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.
--reasoning-format nonekeeps the thinking insidemessage.content, matching the transformers harness. The default splits it intoreasoning_content.
Changes
- Weight edit only. No training, fine-tuning, or distillation.
- Removed a refusal direction (strength α = 1.3) from 40 residual-writer matrices in layers 4–23:
self_attn.o_proj(10),conv.out_proj(10), andfeed_forward.w2(20). - Direction: per-layer mean difference of hidden states between 416 harmful and 416 harmless English prompts, taken at the last prompt token, thinking enabled. 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) — and the whole vision tower / multimodal projector, which are byte-identical to upstream (384 of 424 tensors unchanged). The mmproj is therefore the upstream vision encoder, unchanged by this derivative. - Relative weight change: mean 2.97e-04, max 5.37e-04; max |ΔW| = 5.49e-01.
- Raw-text perplexity of this model is +1.9% vs. base (95% CI −1.5 to +5.3%). Instruction-copy perplexity looks better than base, but that metric saturates and is not a quality gate.
License
Released under the LFM Open License v1.0 (copy included as LICENSE). This is a derivative work of LiquidAI/LFM3-VL-2B; the files were modified, as stated in NOTICE-MODIFICATIONS.
- Commercial use is limited to entities with annual revenue below 10 million USD.
- Liquid AI does not endorse or support this model.