Name: Meta Llama 3 8B Instruct Abliterated V3
Author: failspy

Meta Llama 3 8B Instruct Abliterated V3 is an open-source language model by failspy. Features: 8b LLM, VRAM: 16.1GB, Context: 8K, License: llama3, Instruction-Based, HF Score: 67.3, LLM Explorer Score: 0.26, Arc: 61.6, HellaSwag: 78.4, MMLU: 66.7, TruthfulQA: 52.4, WinoGrande: 74.5, GSM8K: 70.1.

Conversational Endpoints compatible Instruct Llama Region:us Safetensors Sharded Tensorflow

Model Card on HF 🤗: https://huggingface.co/failspy/Meta-Llama-3-8B-Instruct-abliterated-v3

Meta Llama 3 8B Instruct Abliterated V3 Benchmarks

IFEval: 72.45 vs 88 (so35)^-17.7%

ARC: 61.6 vs 96.7 (so35)^-36.3%

HellaSwag: 78.35 vs 95.3 (gpt4)^-17.8%

MMLU: 66.67 vs 88.3 (so35)^-24.5%

TruthfulQA: 52.44 vs 59 (gpt4)^-11.1%

WinoGrande: 74.51 vs 87.5 (gpt4)^-14.8%

GSM8K: 70.05 vs 96.4 (so35)^-27.3%

MATH Lvl 5: 9.59

LLME Score: 0.25883

^nn.n% — How the model compares to the reference models: Anthropic Sonnet 3.5 ("so35"), GPT-4o ("gpt4o") or GPT-4 ("gpt4").

What is the LLM Explorer Rank (Score)

Meta Llama 3 8B Instruct Abliterated V3 (failspy/Meta-Llama-3-8B-Instruct-abliterated-v3)

🌟 Advertise your project 🚀

Meta Llama 3 8B Instruct Abliterated V3 Parameters and Internals

Model Type

text generation

Additional Notes

The model has undergone orthogonalization to remove refusal behavior without altering other aspects of the original 70B instruct model. Known as "abliteration," this process differentiates it from standard uncensored fine-tunes. Potential quirks due to new methodology are acknowledged.

Training Details

Methodology:

Orthogonalization (Ablation) of refusal behavior based on a refined methodology.

LLM Name	Meta Llama 3 8B Instruct Abliterated V3
Repository 🤗	https://huggingface.co/failspy/Meta-Llama-3-8B-Instruct-abliterated-v3
Model Size	8b
Required VRAM	16.1 GB
Updated	2026-04-03
Maintainer	failspy
Model Type	llama
Instruction-Based	Yes
Model Files	5.0 GB: 1-of-4 5.0 GB: 2-of-4 4.9 GB: 3-of-4 1.2 GB: 4-of-4
Model Architecture	LlamaForCausalLM
License	llama3
Context Length	8192
Model Max Length	8192
Transformers Version	4.40.1
Tokenizer Class	PreTrainedTokenizerFast
Vocabulary Size	128256
Torch Data Type	bfloat16

Quantized Models of the Meta Llama 3 8B Instruct Abliterated V3

Model	Likes	Downloads	VRAM
...truct Abliterated V3 GPTQ 4bit	0	12	5 GB

Best Alternatives to Meta Llama 3 8B Instruct Abliterated V3

Best Alternatives	Context / RAM	Downloads	Likes
...otron 8B UltraLong 4M Instruct	4192K / 32.1 GB	89	119
UltraLong Thinking	4192K / 16.1 GB	9	3
...a 3.1 8B UltraLong 4M Instruct	4192K / 32.1 GB	176	24
...a 3.1 8B UltraLong 2M Instruct	2096K / 32.1 GB	875	9
...otron 8B UltraLong 2M Instruct	2096K / 32.1 GB	115	17
Cthulhu 8B V1.4	1048K / 16.1 GB	12	9
...otron 8B UltraLong 1M Instruct	1048K / 32.1 GB	3440	55
...a 3.1 8B UltraLong 1M Instruct	1048K / 32.1 GB	1387	29
Zero Llama 3.1 8B Beta6	1048K / 16.1 GB	7	1
...dger Nu Llama 3.1 8B UltraLong	1048K / 16.2 GB	10	3

Rank the Meta Llama 3 8B Instruct Abliterated V3 Capabilities

🆘 Have you tried this model? Rate its performance. This feedback would greatly assist ML community in identifying the most suitable model for their needs. Your contribution really does make a difference! 🌟

Instruction Following and Task Automation
Factuality and Completeness of Knowledge
Censorship and Alignment
Data Analysis and Insight Generation
Text Generation
Text Summarization and Feature Extraction
Code Generation
Multi-Language Support and Translation

What open-source LLMs or SLMs are you in search of? 53185 in total.

Email us: info@extractum.io. Our Privacy Policy | Terms and Conditions | Suggest an improvement.

Our Social Media →

Original data from HuggingFace, OpenCompass and various public git repos.

Check out Ag3ntum — our secure, self-hosted AI agent for server management.

Release v20260328a

Support LLM Explorer

Meta Llama 3 8B Instruct Abliterated V3 by failspy

» All LLMs » failspy » Meta Llama 3 8B Instruct Abliterated V3 URL Share it on