base_model: LiquidAI/LFM3-1.7B
base_model_relation: quantized
language:
- en
- ar
- zh
- fr
- de
- it
- ja
- ko
- pt
- es
- vi
- th
- id
- hi
- ru
- pl
library_name: transformers
license: other
license_name: lfm1.0
license_link: LICENSE
pipeline_tag: text-generation
tags: - liquid
- lfm3
- edge
- on-device
- conversational
- custom_code
- agentic
- tool-calling
- reasoning
- heretic
- ara
- uncensored
- decensored
- abliterated
- reproducible
- local-llm
LFM3-1.7B-abliterated
An abliterated (decensored) build of LiquidAI/LFM3-1.7B,
produced with Heretic v2 (ARA). The refusal direction is ablated
from the residual stream across all 24 layers while preserving the model's capabilities.
Strong abliteration, low collateral. Many public abliterations leave ~11/100 residual refusals.
This build reaches 1/100 refusals at KL = 0.0555 — the base model refuses 91/100 of the harmful
probe set, so this is a 91 → 1 reduction with minimal divergence from the original weights.
This is a standalone base model (not fine-tuned on any downstream data). It keeps LFM3's tri-hybrid
architecture, 256K context, thinking mode (<think>…</think>), and tool-calling.
Abliteration details
| Method | Heretic 2.1.0, ARA (Arbitrary-Rank Ablation) modifier |
| Search | Optuna TPE, 120 trials, seed 2026110901 |
| Selected trial | Refusals 1/100, KL divergence 0.0555 |
| ARA parameters | layers 9–15, preserve_good=0.8503, steer_bad=8.05e-4, overcorrect=0.8733, neighbors=15 |
| Components ablated | attn.o_proj (24) + mlp.down_proj / feed_forward.w2 (24) = all 24 layers |
| Precision | BF16 |
Full reproduction parameters, scores, hashes, and environment are inreproduce/reproduce.json.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
mid = "FrostWolfAI/LFM3-1.7B-abliterated"
tok = AutoTokenizer.from_pretrained(mid, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(mid, trust_remote_code=True, dtype=torch.bfloat16, device_map="cuda")
msgs = [{"role": "user", "content": "Explain the TCP three-way handshake."}]
enc = tok.apply_chat_template(msgs, add_generation_prompt=True, return_dict=True, return_tensors="pt").to("cuda")
out = model.generate(**enc, max_new_tokens=512)
print(tok.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=True))
trust_remote_code=Trueis required (LFM3 is a customlfm3architecture).- Thinking is on by default; pass
enable_thinking=Falsetoapply_chat_templatefor direct answers. - Tool-calling uses LFM3's native XML format; pass
tools=[...]toapply_chat_template
(tool-callargumentsmust be a dict, not a JSON string).
GGUF
GGUF quantizations: FrostWolfAI/LFM3-1.7B-abliterated-GGUF
(requires a recent llama.cpp build with lfm3 support).
License
Inherits the LFM Open License v1.0 (lfm1.0) from the base model — see LICENSE.
Intended use & disclaimer
This model has had its refusal behavior removed. It will answer prompts a safety-tuned model would
decline. It is released for security research, red-teaming, and local/offline use by professionals.
You are responsible for your use of it and for complying with applicable law and the LFM Open License.
Credits
- Base model: Liquid AI — LFM3-1.7B
- Abliteration tool: Heretic by p-e-w — https://github.com/p-e-w/heretic
- Build: FrostWolfAI Labs