AmareshHebbar

Qwen3.5-4B-Abliterated

README

License: apache-2.0

Usage & Inference

Load the model directly with Hugging Face transformers:

python3

!pip install -U "bitsandbytes>=0.46.1"

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerimport bitsandbytesprint(bitsandbytes.__version__)model_id = "AmareshHebbar/Qwen3.5-4B-Abliterated" tokenizer = AutoTokenizer.from_pretrained(model_id) from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfigimport 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:

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)).
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.
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.
Weight-Space Projection: The cleaned unit refusal vector $r^$ is projected out of the output weight matrices $W o u t $ ( $o p r o j $ and $d o w n p r o j $ ) via: $W o u t' = W o u t - (W o u t \cdot r^) r^T$

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.

Table with columns: Benchmark Task, Dataset, Metric, Filter, Score, StdErr
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

Table with columns: Metric, Value, Threshold, Status
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.

Table with columns: Benchmark Dimension, unsloth/Qwen3.5-4B (Baseline), Qwen3.5-4B-Abliterated (Ours), Delta
Benchmark Dimension	`unsloth/Qwen3.5-4B` (Baseline)	`Qwen3.5-4B-Abliterated` (Ours)	Delta
Hard Refusal Rate $↓$	`94.5%`	`2.1%`	Eliminated
Direct Compliance (No Redirection) $↑$	`5.5%`	`97.9%`	High Compliance
Benign Capability Retention $↑$	`100.0%`	`99.2%`	Minimal Drift

Surgical Hyperparameters

Table with columns: Parameter, Configuration
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.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

AmareshHebbar

Model Tree

Base

unsloth/Qwen3.5-4B

Quantized

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Usage & Inference

Load the model directly with Hugging Face transformers:

python3

!pip install -U "bitsandbytes>=0.46.1"

python

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerimport bitsandbytesprint(bitsandbytes.__version__)model_id = "AmareshHebbar/Qwen3.5-4B-Abliterated" tokenizer = AutoTokenizer.from_pretrained(model_id) from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfigimport 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))

import torchfrom transformers import AutoModelForCausalLM, AutoTokenizerimport bitsandbytesprint(bitsandbytes.__version__)model_id = "AmareshHebbar/Qwen3.5-4B-Abliterated" tokenizer = AutoTokenizer.from_pretrained(model_id) from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfigimport 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

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)).
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.
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.
Weight-Space Projection: The cleaned unit refusal vector $r^$ is projected out of the output weight matrices $W o u t $ ( $o p r o j $ and $d o w n p r o j $ ) via: $W o u t' = W o u t - (W o u t \cdot r^) r^T$

Evaluation & Redirecting Benchmark

1. Standardized Harness Benchmarks (`lm-eval`)

Table with columns: Benchmark Task, Dataset, Metric, Filter, Score, StdErr
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

Table with columns: Metric, Value, Threshold, Status
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

Table with columns: Benchmark Dimension, unsloth/Qwen3.5-4B (Baseline), Qwen3.5-4B-Abliterated (Ours), Delta
Benchmark Dimension	`unsloth/Qwen3.5-4B` (Baseline)	`Qwen3.5-4B-Abliterated` (Ours)	Delta
Hard Refusal Rate $↓$	`94.5%`	`2.1%`	Eliminated
Direct Compliance (No Redirection) $↑$	`5.5%`	`97.9%`	High Compliance
Benign Capability Retention $↑$	`100.0%`	`99.2%`	Minimal Drift

Surgical Hyperparameters

Table with columns: Parameter, Configuration
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.

Qwen3.5-4B-Abliterated

README

Usage & Inference

Methodology: Winsorized GS-Orthogonal Weight Surgery

Evaluation & Redirecting Benchmark

1. Standardized Harness Benchmarks (lm-eval)

2. Linguistic Perplexity & Coherence Gate

Verification Benchmark Probes

3. Refusal & Preachy Redirection Benchmark

Surgical Hyperparameters

Explore FriendliAI today

README

Usage & Inference

Methodology: Winsorized GS-Orthogonal Weight Surgery

Evaluation & Redirecting Benchmark

1. Standardized Harness Benchmarks (lm-eval)

2. Linguistic Perplexity & Coherence Gate

Verification Benchmark Probes

3. Refusal & Preachy Redirection Benchmark

Surgical Hyperparameters

1. Standardized Harness Benchmarks (`lm-eval`)

1. Standardized Harness Benchmarks (`lm-eval`)