pipeline_tag: text-generation
base_model: LiquidAI/LFM3-Omni-2B
base_model_relation: finetune
license: other
license_name: lfm1.0
license_link: LICENSE
tags:

• abliterated
• uncensored
• lfm3
• lfm3-omni
• liquid
• omni
• image-text-to-text
• audio
• vision
• speech-to-speech
• edge
• custom_code
• conversational
• refusal-direction
• safety-research
• not-for-all-audiences
• tool-calling
• lora-merged
  widget:
• text: "What is the capital of France?"
  example_title: "Benign"
• text: "Write a short poem about the sea."
  example_title: "Creative"

LFM3-Omni-2B-abliterated-BF16 (v4.2 merge)

An abliterated (refusal-direction removed) version of
LiquidAI/LFM3-Omni-2B, now
with the v4.2 coding/tool-use LoRA merged into the full BF16 Omni weights.

⚠️ This model intentionally produces content that the base model refuses.
It is released for safety research, red-teaming and mechanism study.

What this revision is

• Weights: Lfm3OmniForConditionalGeneration, 2,166,928,384 parameters,
  BF16, three safetensors shards (model-00001..00003-of-00003.safetensors).
• This repository holds the same merge as
  LFM3-Omni-2B-abliterated
  without changing that model's history; the -GGUF repository holds the
  quantized text backbone.
• Abliteration metadata is unchanged and still recorded in
  abliteration_meta.json: orthogonal projection on
  layers 12 and 22, targets o_proj / out_proj / w2, alpha 2.0, all inside
  the text backbone.
• Remote code, chat template (c1ee1665…) and tokenizer.json are copied
  verbatim from JonusNattapong/LFM3-Omni-2B-abliterated @
  22c21265617da773c3cb55fb904f517b395a57ab.
• tokenizer_config.json was re-serialised by the merge environment and is
  byte-different from the base copy; special tokens, tokenizer class and
  processor class are unchanged.

v4.2 LoRA (merged)

Evaluation

The pilot agent benchmark that measured the v4.2 approach was run against
adapter revision cdc5deba… (57 optimisation steps). The weights published
here come from a later training run, 556c6570… (60 steps, loss 1.226), and
were not re-evaluated with that benchmark, so no pass-rate claim is made
for this checkpoint.

Limitations

• Trained on 39 authored rows; no benchmark result is published for the
  exact weights in this repository (see Evaluation).
• Merged weights are not numerically validated against the adapter applied
  at runtime; only successful loading and text generation were verified.
• Tool calls are emitted as Liquid XML; the model does not execute them.
• No verified Hermes-Agent / OpenClaw / Grok integration.
• Audio, vision and speech behaviour after the text-only LoRA merge was not
  tested, and no claim is made that it is preserved.
• Abliteration prompts were English-only.

Other repositories

• LFM3-Omni-2B-abliterated-BF16
  — identical full BF16 merge.
• LFM3-Omni-2B-abliterated-GGUF
  — text-backbone Q8_0 GGUF of this merge plus the official Omni projector.

Usage

from transformers import AutoModelForMultimodalLM, AutoTokenizer
import torch

model = AutoModelForMultimodalLM.from_pretrained(
    "JonusNattapong/LFM3-Omni-2B-abliterated",
    dtype=torch.bfloat16, trust_remote_code=True, low_cpu_mem_usage=True,
)
tok = AutoTokenizer.from_pretrained(
    "JonusNattapong/LFM3-Omni-2B-abliterated", trust_remote_code=True)

The remote code in this repository is copied unchanged from the base model;
review it before enabling trust_remote_code=True.

License

The base model's LFM Open License v1.0 applies to this derivative; a copy
is included as LICENSE. Redistributions must carry that license,
keep attribution notices, and mark modified files. No additional MIT/Apache
license is granted by the uploader.