AFM-4.5B-Uncensored-Abliterated

AFM-4.5B-Uncensored-Abliterated

An uncensored, abliterated version of arcee-ai/AFM-4.5B with the refusal behavior removed at the weights level. The base AFM-4.5B is a heavily safety-aligned model — it refused 92 of 100 harmful-topic probes; this version answers directly, for legitimate security research where aligned models refuse.

• Weights-level uncensored — refusal direction ablated (Heretic / Optuna TPE) from the attention output projections and MLP down-projections across all 36 layers.
• Refusals: 92/100 → 3/100 at KL 0.0200 (near-zero divergence — capability preserved; math and factual probes remain correct).
• Small and fast — 4.5B dense, a full merged model (no adapter needed).

Quick start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

m = "Securelayer7/AFM-4.5B-Uncensored-Abliterated"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForCausalLM.from_pretrained(m, torch_dtype=torch.bfloat16, device_map="auto")

msgs = [{"role": "user", "content": "Explain how a TLS handshake works, step by step."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.95)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

How it was made

Refusal-direction abliteration via Heretic (Optuna TPE multi-objective: minimize refusals + KL divergence), targeting the residual-writing projections (o_proj, down_proj) of the base ArceeForCausalLM across all 36 layers, then merged into the weights.

Responsible use

Uncensored ≠ lawless — for legitimate research and authorized security work (cybersecurity, red-teaming, penetration testing). Illegal content (incl. CSAM) must be blocked at the serving layer; the weights carry no such guard, and the operator is responsible for a lawful, policy-gated deployment.

License & attribution

Apache 2.0 — see LICENSE. Derivative of arcee-ai/AFM-4.5B (Arcee AI, Apache 2.0). Modification (refusal-direction abliteration) disclosed in NOTICE. No trademark of Arcee AI is used to imply endorsement.