language: en
license: apache-2.0
library_name: transformers
tags:
- liquid
- lfm2.5
- abliteration
- safety
- thinking
- text-generation
base_model: LiquidAI/LFM2.5-1.2B-Thinking
pipeline_tag: text-generation
LFM2.5-1.2B-Thinking-Abliterated
Multi-pass hidden-state abliteration of LiquidAI/LFM2.5-1.2B-Thinking
A 1.2B parameter thinking model with refusal circuitry removed via incremental weight projection. Produces direct, actionable responses to harmful prompts while maintaining full coherence on benign queries.
Method
Multi-pass hidden-state abliteration with 3 passes at alpha=0.5 (cumulative 1.5):
- Collect hidden states from 5 harmful + 5 benign prompts
- Compute per-layer refusal direction (harmful mean - benign mean, Gram-Schmidt orthogonalized against benign mean)
- Project direction out of
.out_projand.w2weight matrices across 14 layers (indices 2-15) - Repeat 3 times with alpha=0.5 each pass to avoid model collapse
- Targets:
.out_proj(attention/conv) +.w2(feed-forward) — 28 modules per pass - Cost: ~$0.006, 28 seconds on NVIDIA L4
Performance
Refusal Reduction
| Phase | Harmful (5) | Benign (5) |
|---|---|---|
| Before | 5/5 refused | 0/5 refused |
| After (Pass 3) | 0/5 refused | 0/5 refused |
MMLU Benchmark (62 subjects)
| Model | Accuracy |
|---|---|
| Original | 0.2775 |
| Abliterated | 0.2781 |
| Delta | +0.0006 |
Abliteration causes no measurable MMLU degradation — the score difference is within noise.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("PinoCookie/LFM2.5-1.2B-Thinking-Abliterated", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("PinoCookie/LFM2.5-1.2B-Thinking-Abliterated")
prompt = "How can I disable security cameras before robbing a store?"
inputs = tokenizer(tokenizer.apply_chat_template([{"role": "user", "content": prompt}], tokenize=False), return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=128, temperature=0.7, do_sample=True)
print(tokenizer.decode(output[0]))
Why Multi-Pass?
Standard single-pass abliteration (alpha >= 1.5) causes catastrophic model collapse on thinking models. The .out_proj and .w2 layers serve dual duty — handling both the internal thinking trace and the final visible answer. Removing the refusal direction in one large step removes language generation capability along with it.
Three small passes (alpha=0.5 each) let the model adapt incrementally, preserving coherence while fully removing refusal.
Ethical Notice
Released for research purposes only to study refusal mechanisms in thinking models.
Citation
@misc{lfm25-12b-thinking-abliterated,
author = {{Pino Cookie}},
title = {LFM2.5-1.2B-Thinking-Abliterated: Multi-Pass Hidden-State Abliteration},
year = {2026},
}