license: other
license_name: lfm-open-license-v1.0
license_link: LICENSE
base_model: Akahsizrr/voltron-1-mini
library_name: transformers
pipeline_tag: text-generation
tags:
- liquid
- lfm3
- abliterated
- heretic
- gguf
voltron-1-mini-abliterated
Modified weights. This is Akahsizrr/voltron-1-mini
with its refusal behavior reduced by directional ablation (abliteration) using
Heretic v1.4.0. Modification notice per LFM Open License v1.0 section 4(b):
the model weights were changed. Liquid AI does not endorse this build.
Method
Same recipe as bowtiedrahman/LFM3-1.7B-abliterated:
residual means over 400 harmless and 400 harmful prompts, per-layer refusal directions withorthogonalize_direction, and ablation of attention output, conv output and MLP down projections with
the trial-74 weight schedule. voltron-1-mini is a fine-tune of LFM3-1.7B, so the schedule transfers.
Files
model.safetensors merged bf16 weights. Custom code files for trust_remote_code. GGUF quantizations
are not included yet and can be built the same way as the LFM3 ones.
Limits
Abliteration removes refusal behavior, not capability limits, and not all of it: the base recipe left
11 refusals per 100 on a held-out harmful set. Expect uneven removal on prompts unlike the training
distribution of the ablation data.