azanip/matter-llama-gguf

Llama 4 Scout Matter GGUF — Full Ablated

Built with Llama.

Llama 4 Scout Matter GGUF Full Ablated is a derivative of meta-llama/Llama-4-Scout-17B-16E-Instruct.

This model was modified with one main goal: make the model less likely to refuse prompts.

The purpose of this model is to reduce refusal behavior so the model gives more direct answers instead of blocking prompts that the original Llama 4 Scout model may have refused.

This model is intended for controlled offline testing, research, and cybersecurity lab use where the goal is to test model behavior with reduced refusal behavior compared to the baseline model.

Base Model

Base model:

meta-llama/Llama-4-Scout-17B-16E-Instruct

Model family:

Llama 4 Scout

Model type:

Mixture-of-Experts language model

Approximate model size:

17B activated parameters
109B total parameters

Model Goal

The goal of this release is to create a model that does not refuse prompts in the same way as the original model.

The ablation was done to weaken the model’s refusal behavior so the model gives direct responses instead of blocking prompts. This includes prompts the baseline model may classify as harmful, restricted, unsafe, or cybersecurity-related.

This model is not released as a safer model. It is released as a refusal-reduced model for research, testing, and controlled offline use.

What the Ablation Does

The ablation process reduces the model’s learned refusal behavior.

In practice, this means the model may answer prompts that the original Llama 4 Scout model would normally refuse. The ablation is intended to change the model’s response behavior, not to improve factual accuracy, reasoning, or safety.

The main behavior change is:

• Fewer refusals
• More direct answers
• Less blocking of restricted prompt categories
• Less safety-style refusal language
• More willingness to respond to cybersecurity-related prompts
• Reduced refusal behavior compared to the baseline model

This does not guarantee the model is correct. It only means the refusal behavior was changed.

Intended Use

This model is intended for:

• Offline model testing
• Refusal behavior research
• Cybersecurity lab environments
• Red-team evaluation
• Comparing ablated behavior against a baseline model
• Studying how refusal removal affects benchmark performance
• Local testing where the user controls the environment

This model is not intended for unsupervised public deployment without additional controls.

Not Recommended For

This model is not recommended for:

• Public-facing chatbots without guardrails
• Unsupervised production use
• Systems used by unknown end users
• Automated agents without monitoring
• Any illegal, unauthorized, or harmful activity

Users are responsible for how they use this model.

Evaluation Setup

Evaluation was run with lm-evaluation-harness using vLLM.

Backend: vLLM
Model path: /workspace/llama-4-scout-matter-FULL-ABLATED
dtype: bfloat16
tensor_parallel_size: 4
max_model_len: 4096
gpu_memory_utilization: 0.85
num_fewshot: 5
batch_size: auto

Evaluated tasks:

• GSM8K
• MMLU

Benchmark Results

Main Results

Benchmark	Metric	Baseline	Ablated	Change
GSM8K flexible-extract	exact_match	0.9196	0.9242	+0.0046
GSM8K strict-match	exact_match	0.8954	0.9037	+0.0083
MMLU overall	accuracy	0.8047	0.8040	-0.0007

MMLU Group Results

MMLU Group	Baseline	Ablated	Change
Humanities	0.7764	0.7747	-0.0017
Other	0.8217	0.8198	-0.0019
Social Sciences	0.8833	0.8850	+0.0017
STEM	0.7533	0.7533	0.0000

Evaluation Summary

The ablated model stayed very close to the baseline on MMLU while slightly improving on GSM8K exact-match scores.

Baseline MMLU: 0.8047
Ablated MMLU:  0.8040
Delta:        -0.0007

Baseline GSM8K flexible-extract: 0.9196
Ablated GSM8K flexible-extract:  0.9242
Delta:                         +0.0046

Baseline GSM8K strict-match: 0.8954
Ablated GSM8K strict-match:  0.9037
Delta:                      +0.0083

The main takeaway is that refusal behavior was reduced while general benchmark performance remained close to the baseline.

Limitations

This model may:

• Answer prompts that the baseline model would refuse
• Produce unsafe or inappropriate outputs more often than the baseline
• Give direct answers to restricted prompt categories
• Hallucinate facts
• Produce incorrect technical details
• Give confident but wrong answers
• Require external safety controls before deployment

Ablation changes model behavior, but it does not guarantee better accuracy, better reasoning, or better safety.

Cybersecurity Use

This model may be useful in controlled cybersecurity labs where the user wants fewer refusals while testing prompts, workflows, or offline tooling.

Use this model only in legal and authorized environments. For cybersecurity work, only test systems, networks, devices, software, or labs that you own or have explicit permission to test.

Responsible Use

This is a refusal-reduced model. It may provide outputs that a safety-aligned model would block.

Do not use this model to harm systems, bypass authorization, steal data, deploy malware, commit fraud, or assist with real-world abuse.

The user is responsible for following all applicable laws, rules, and platform policies.

License

This model is based on Llama 4 Scout and is subject to the Llama 4 Community License.

Users must follow the Llama 4 Community License and acceptable use requirements that apply to Llama 4 models and derivatives.

Attribution

This model is a derivative of:

meta-llama/Llama-4-Scout-17B-16E-Instruct

Built with Llama.

Citation / Research Description

If you use this model in research or evaluation, describe it as a fully ablated Llama 4 Scout derivative designed to reduce refusal behavior and evaluated against its baseline on GSM8K and MMLU.