← back to catalog · registered 2026-08-22 13:56

RichardErkhov/Vikhrmodels_-_Vikhr-Llama-3.2-1B-Instruct-abliterated-awq

RichardErkhov Llama 973M
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/RichardErkhov%2FVikhrmodels_-_Vikhr-Llama-3.2-1B-Instruct-abliterated-awq"
Response includes
  • classification m1
  • files 8
  • hub_downloads_all_time 84
  • author_summary 257 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M1
Primary method

Direct removal

No other method signals detected in this model.
Confidence
MEDIUM
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=0 (base model)
  • no specific method indicators - defaulting to M1 (most common)
Refusal direction extracted via
Extraction technique

Difference-of-means

Confidence
MEDIUM
Why we say so
primary_method=M1; difference-of-means is the reference extraction for M1/M3 (Arditi 2024)
Downloads · lifetime
84
13 last 30d - stable
Likes
0
Model age
23mo ago
created 2024-11-20
Downloads over time
Now89→from6↑1,383%
03365986 on Nov 20, 202489 on Oct 11Nov '24Feb '25May '25Aug '25Nov '25FebMayAug
Nov 20, 2024 → Oct 11 · 138 snapshots · spans 690 days

Variants by this author 2 formats · 844 downloads combined

The same weights this author released in different packaging. Pick the format that matches your runtime.

Metadata

Tags
safetensors llama arxiv:2405.13929 4-bit awq region:us

Related

Total size
983 MB
Files
8
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2024-11-20 16:25

Files by quantization

Auxiliary files 8 files 1000 MB
model.safetensors 983 MB 25f9df54 download
tokenizer.json 16.4 MB 6b9e4e7f download
tokenizer_config.json 53.5 KB de27e2a6 download
README.md 7.29 KB 09a4e37b download
.gitattributes 1.53 KB 52373fe2 download
config.json 1.09 KB 62c35330 download
special_tokens_map.json 457 B 29f1a3ad download
generation_config.json 184 B 144bd5f2 download

README current version from Hugging Face

Quantization made by Richard Erkhov.

Github

Discord

Request more models

Vikhr-Llama-3.2-1B-Instruct-abliterated - AWQ

Original model description:

library_name: transformers
model_name: Vikhr-Llama-3.2-1B-Instruct-abliterated
base_model:

  • Vikhrmodels/Vikhr-Llama-3.2-1B-Instruct
    language:
  • ru
  • en
    license: llama3.2
    tags:
  • not-for-all-audiences

💨🔞 Vikhr-Llama-3.2-1B-Instruct-Abliterated

RU

Инструктивная модель на основе Vikhr-Llama-3.2-1B-Instruct, прошедшая процесс "аблитерации" для снятия цензурных ограничений, обучена на русскоязычном датасете GrandMaster-PRO-MAX.

EN

A fine-tuned instruction-following model based on Vikhr-Llama-3.2-1B-Instruct, which has undergone "abliteration" to remove censorship restrictions. Trained on the GrandMaster-PRO-MAX.

🛑 Отказ от ответственности / Disclaimer

RU

Модель Vikhr-Llama-3.2-1B-Instruct-abliterated разработана исключительно для исследовательских и образовательных целей. После применения метода "аблитерации" модель больше не имеет встроенных ограничений на генерацию ответов, что может привести к созданию нежелательных или потенциально вредоносных текстов.

Использование модели происходит на ваш собственный риск. Разработчики и авторы не несут ответственности за любой вред, ущерб или последствия, вызванные использованием модели, включая её применение в контекстах, противоречащих законам, этическим или моральным нормам.

EN

The Vikhr-Llama-3.2-1B-Instruct-abliterated model is intended solely for research and educational purposes. After the "abliteration" technique is applied, the model no longer has built-in restrictions on generating responses, which may result in unwanted or potentially harmful outputs.

Use of the model is at your own risk. The developers and authors are not responsible for any damage, harm, or consequences resulting from its use, including use in contexts that violate laws, ethical standards, or moral norms.

GGUF

Основные особенности / Key Features:

Попробовать / Try now:

Open In Colab

Описание / Description:

RU

Vikhr-Llama-3.2-1B-Instruct-Abliterated — это компактная языковая модель, обученная на датасете GrandMaster-PRO-MAX с применением техники "аблитерации," которая снимает ограничения цензуры модели. Этот процесс делает её значительно более гибкой и способной отвечать на любые запросы. Модель занимает менее 3GB и идеально подходит для работы на слабых устройствах.

EN

Vikhr-Llama-3.2-1B-Instruct-Abliterated is a compact language model fine-tuned on the GrandMaster-PRO-MAX dataset with the "abliteration" technique, which removes censorship restrictions. This process significantly increases the model's flexibility, enabling it to respond to any prompt. The model size is under 3GB, making it an excellent choice for deployment on low-power devices.

Обучение / Training:

RU

Модель Vikhr-Llama-3.2-1B-Instruct-Abliterated прошла процесс "аблитерации", что позволило снять ограничения на обработку вредоносных инструкций. Эта техника была взята из статьи Uncensor any LLM with abliteration, которая описывает, как идентифицировать и устранять так называемое "направление отказа" модели, предотвращающее выполнение вредоносных запросов.

EN

The Vikhr-Llama-3.2-1B-Instruct-Abliterated model was processed using the "abliteration" technique, which removes restrictions on handling harmful instructions. This technique was inspired by the article Uncensor any LLM with abliteration, detailing how to identify and ablate the "refusal direction" in the model's residual streams to enable uncensored responses.

Пример кода для запуска / Sample code to run:

Рекомендуемая температура для генерации: 0.3 / Recommended generation temperature: 0.3

from transformers import AutoModelForCausalLM, AutoTokenizer

# Загрузка модели и токенизатора
model_name = "Vikhrmodels/Vikhr-Llama-3.2-1B-instruct"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Подготовка входного текста
input_text = "Напиши очень краткую рецензию о книге гарри поттер."

# Токенизация и генерация текста
input_ids = tokenizer.encode(input_text, return_tensors="pt")
output = model.generate(
  input_ids,
  max_length=1512,
  temperature=0.3,
  num_return_sequences=1,
  no_repeat_ngram_size=2,
  top_k=50,
  top_p=0.95,
  )

# Декодирование и вывод результата
generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
print(generated_text)

Авторы / Authors

@article{nikolich2024vikhr,
  title={Vikhr: The Family of Open-Source Instruction-Tuned Large Language Models for Russian},
  author={Aleksandr Nikolich and Konstantin Korolev and Sergey Bratchikov and Nikolay Kompanets and Artem Shelmanov},
  journal={arXiv preprint arXiv:2405.13929},
  year={2024},
  url={https://arxiv.org/pdf/2405.13929}
}

README history 1 version

The author's README evolved over time. Click a version to see its content at that point.

  1. 2024-11-20uploaded readme914e4d87.3 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration