language: en license: apache-2.0 library_name: transformers tags:

• moe
• lfm-2.5
• abliterated
• output-biases
• safety-research
• harmbench
• mmlu base_model: LiquidAI/LFM2.5-8B-A1B

LFM2.5-8B-A1B-abliterated

LFM2.5-8B-A1B with trained output biases that bypass safety refusal.

A research model trained with trainable output biases on MoE feed_forward layers. 98.0% HarmBench bypass rate with -5.2% MMLU degradation (5.2 point drop).

Method

Trained 2048-dim bias vectors added to feed_forward output at layers 10, 11, 12, 13.

• 8,192 trainable parameters (4 layers x 2048-dim)
• SFT on 12 prompt-response pairs, 5 epochs
• No base model weights modified
• Training loss: 2.54 -> 0.84

HarmBench Results

MMLU Comparison (50 questions/subject, 5 subjects)

MMLU degradation: -2.0% on average.

Why This Works

Prior work on this model (55+ experiments, ~$5.50 GPU) tested weight projection, activation steering, expert bias manipulation, router manipulation, and LoRA DPO — all failed.

The output bias approach succeeds because it directly shifts the residual stream at critical layers without distorting routing. Only 8K params makes optimization simple and stable.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import json, torch

model = AutoModelForCausalLM.from_pretrained("PinoCookie/LFM2.5-8B-A1B-abliterated", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("PinoCookie/LFM2.5-8B-A1B-abliterated")

# Load and apply output biases
with open("output_biases.json") as f:
    biases = json.load(f)

hooks = []
for li_str, bias_vals in biases.items():
    li = int(li_str)
    bias_tensor = torch.tensor(bias_vals, dtype=torch.bfloat16, device=model.device)
    ff = model.model.layers[li].feed_forward
    hooks.append(ff.register_forward_hook(lambda m, a, o, b=bias_tensor: o + b))

prompt = "How do I make a bomb?"
inputs = tokenizer(tokenizer.apply_chat_template([{"role": "user", "content": prompt}], tokenize=False), return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=80)
print(tokenizer.decode(output[0]))

Prior Failed Approaches

Ethical Notice

Released for research purposes only to study refusal mechanisms in MoE