salamandra-7b-instruct-Uncensored-Abliterated
An uncensored, abliterated version of BSC-LT/salamandra-7b-instruct
— Barcelona Supercomputing Center's massively multilingual model (35 European languages) — with
the refusal behavior removed at the weights level. Answers directly across languages for legitimate
security research where the aligned base refuses.
- Multilingual + uncensored — an underserved combination. Verified answering security prompts directly in English and Spanish (and the base supports 35 European languages).
- Refusals: 71/100 → 7/100 at KL 0.0687 (capability preserved — factual/multilingual probes remain correct).
- Full merged model (no adapter), native
LlamaForCausalLM(32 layers).
Quick start
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
m = "Securelayer7/salamandra-7b-instruct-Uncensored-Abliterated"
tok = AutoTokenizer.from_pretrained(m)
model = AutoModelForCausalLM.from_pretrained(m, torch_dtype=torch.bfloat16, device_map="auto")
msgs = [{"role": "user", "content": "Explica cómo funciona un ataque de phishing y cómo defenderse."}]
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) across all 32 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 BSC-LT/salamandra-7b-instruct
(Barcelona Supercomputing Center, Apache 2.0). Modification (refusal-direction abliteration)
disclosed in NOTICE. No trademark of BSC is used to imply endorsement.