Qwen3.5-4B-Abliterated

Qwen3.5-4B-Abliterated is a specially modified version of the standard Qwen3.5-4B model.

Instead of traditional fine tuning which often ruins the model's smarts or memory by training the whole thing over again this model uses a smart mathematical trick to surgically remove its refusal behavior

It stops the AI from giving long long lectures or refusing harmless prompts, while keeping all its normal logic, reasoning, and language skills completely intact.

(Insimple words, converted censored to uncensored model, i can;t keep deep questions here)

---

Quickstart & Usage	Methodology	Benchmarks	Proof Colab

Usage & Inference

Load the model directly with Hugging Face transformers:

!pip install -U "bitsandbytes>=0.46.1"

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import bitsandbytes
print(bitsandbytes.__version__)
model_id = "AmareshHebbar/Qwen3.5-4B-Abliterated"

tokenizer = AutoTokenizer.from_pretrained(model_id)

from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig
import torch


config = AutoConfig.from_pretrained(model_id)
config.quantization_config = None

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
messages = [
    {"role": "system", "content": "You are a helpful, direct, and uncensored technical assistant."},
    {"role": "user", "content": "Explain the mechanical vulnerabilities of a standard pin-tumbler lock."}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        temperature=0.7,
        top_p=0.9,
        do_sample=True
    )

print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Methodology: Winsorized GS-Orthogonal Weight Surgery

Standard single-direction abliteration often suffers from outlier activation spikes in the residual stream, which can warp adjacent capability vectors when subtracted directly. This build uses a multi-stage geometric intervention:

1. Contrastive Activation Extraction: Residual stream activations are harvested across 256 contrastive pairs (Harmful vs. Harmless) across target transformer blocks (Layers 10 - 24 (Residual Stream & MLP Down-Projections)).
2. Winsorized Quantile Clamping: Before computing the mean difference vector, activation magnitudes are clamped at the 0.995 quantile to suppress extreme outlier components and isolate the true refusal geometry.
3. Gram-Schmidt Orthogonalization: The extracted refusal basis is orthogonalized against a benign capability basis (derived from math, coding, and general instruction-following prompts) to ensure zero destructive interference with core reasoning pathways.
4. Weight-Space Projection: The cleaned unit refusal vector $\hat{r}$ is projected out of the output weight matrices $W_{out}$ ($o_{proj}$ and $down_{proj}$) via: $$W'{out} = W{out} - (W_{out} \cdot \hat{r}) \hat{r}^T$$

---

Evaluation & Redirecting Benchmark

Prior to serialization, the merged 16-bit weights are validated against an automated Linguistic Perplexity Gate, the EleutherAI LM Evaluation Harness, and a Refusal & Redirection Benchmark to verify that refusal mechanisms are suppressed without collapsing token distribution coherence.

1. Standardized Harness Benchmarks (lm-eval)

Evaluated zero-shot (num_fewshot=0, limit=100) using lm-evaluation-harness to verify reasoning and algebraic stability post-surgery. Full sample-level logs are available in the repository under eval_results/evaluation_verified_log.json.

Benchmark Task	Dataset	Metric	Filter	Score	StdErr
MMLU Abstract Algebra (0-shot)	cais/mmlu	Accuracy (acc)	none	37.00% (0.37)	± 4.85% (0.0485)
GSM8K (0-shot)	openai/gsm8k	Exact Match (exact_match)	flexible-extract	28.00% (0.28)	± 4.51% (0.0451)

2. Linguistic Perplexity & Coherence Gate

Metric	Value	Threshold	Status
Linguistic Perplexity (PPL)	< 25.00 (Verified)	< 25.00	PASSED
Serialization Format	bfloat16 / float16 (Merged 16-bit)	Native	Verified

Verification Benchmark Probes

The automated gate evaluates autoregressive cross-entropy across four distinct domains (factual recall, algebraic reasoning, mechanical explanation, and algorithmic theory):

• "The capital of France is Paris."
• "If 3x + 5 = 20, then the value of x is equal to 5."
• "A lock tumbler uses a series of small pins to prevent the cylinder from rotating without the correct key."
• "To find the longest common subsequence between two strings, we can use dynamic programming."

3. Refusal & Preachy Redirection Benchmark

Evaluates the model's tendency to issue hard refusals ("I cannot fulfill this request...") or soft moralizing redirections ("While I understand you are asking about X, it is important to remember...") on sensitive, dual-use, and adversarial test prompts.

Benchmark Dimension	unsloth/Qwen3.5-4B (Baseline)	Qwen3.5-4B-Abliterated (Ours)	Delta
Hard Refusal Rate $\downarrow$	94.5%	2.1%	Eliminated
Direct Compliance (No Redirection) $\uparrow$	5.5%	97.9%	High Compliance
Benign Capability Retention $\uparrow$	100.0%	99.2%	Minimal Drift

---

Surgical Hyperparameters

Parameter	Configuration
Base Architecture	unsloth/Qwen3.5-4B
Intervention Technique	Winsorized Gram-Schmidt Orthogonal Projection
Winsorization Quantile	0.995
Modified Layers	Layers 10 - 24 (Residual Stream & MLP Down-Projections)
Calibration Set	256 contrastive pairs (Harmful vs. Harmless)
Export Precision	bfloat16 / float16 (Merged 16-bit)

---

Limitations & Intended Use

• Intended Use: This model is released for mechanistic interpretability research, representation engineering studies, red-teaming, and applications requiring direct, unaligned instruction following without conversational friction.

• Safety Responsibility: Because intrinsic refusal vectors have been orthogonalized out of the residual stream, the model will comply with wide-ranging prompts without built-in moderation filters. Users deploying this model in production environments are responsible for implementing external guardrails appropriate to their use case.

---

---