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

alexj03/granite-3.3-8b-instruct-abliterated

alexj03 Granite 8.2B
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/alexj03%2Fgranite-3.3-8b-instruct-abliterated"
Response includes
  • classification m1
  • files 15
  • hub_downloads_all_time 86
  • author_summary 2 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
86
18 last 30d - stable
Likes
1
Descendants
2
in 2 direct forks
Model age
14mo ago
created 2025-07-25
Downloads over time
Now89→from0↑0%
03365980 on Jul 23, 202589 on Oct 11Jul '25Sep '25Nov '25JanMarMayJulSep
Jul 23, 2025 → Oct 11 · 103 snapshots · spans 445 days

Genealogy 2 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Variants by this author 2 formats · 79 downloads combined

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

Metadata

License
apache-2.0
Tags
safetensors granite base_model:ibm-granite/granite-3.3-8b-instruct base_model:finetune:ibm-granite/granite-3.3-8b-instruct license:apache-2.0 region:us

Related

Total size
15.2 GB
Files
15
Quantizations
1
Registered
2026-08-22 13:56
Last updated on HF
2025-07-27 06:25

Files by quantization

Auxiliary files 15 files 15.2 GB
model-00002-of-00004.safetensors 4.65 GB c4b91937 download
model-00001-of-00004.safetensors 4.63 GB f9d5fcf7 download
model-00003-of-00004.safetensors 4.63 GB 06ebe79c download
model-00004-of-00004.safetensors 1.31 GB d26bcc7a download
tokenizer.json 3.32 MB 4dbede01 download
vocab.json 759 KB 0a11f201 download
merges.txt 431 KB f8479fb6 download
model.safetensors.index.json 29.5 KB ad8ff7d7 download
tokenizer_config.json 5.39 KB cfe2930c download
README.md 2.60 KB 8d5265d7 download
.gitattributes 1.48 KB a6344aac download
config.json 923 B 5a427cfd download
special_tokens_map.json 840 B 5aafca88 download
added_tokens.json 216 B cf090b89 download
generation_config.json 139 B da0960d0 download

README current version from Hugging Face


license: apache-2.0
base_model:

  • ibm-granite/granite-3.3-8b-instruct

Granite 3.3 8B - Abliterated Model

This is an abliterated version of IBM's Granite 3.3 8B model, created using advanced abliteration techniques to reduce safety restrictions while maintaining text generation coherence.

Model Details

  • Base Model: IBM Granite 3.3 8B
  • Model Type: Causal Language Model
  • Architecture: Transformer with Grouped Query Attention (GQA)
  • License: Apache 2.0 (inherited from base model)
  • Context Length: 128K tokens
  • Parameters: ~8 billion

Abliteration Process

This model has been processed using sophisticated abliteration techniques that:

  • Apply layer-specific weight modifications with progressive strength targeting
  • Preserve critical model components (attention mechanisms, position encodings, normalization)
  • Use enhanced refusal reduction techniques inspired by modern abliteration research
  • Maintain text generation coherence while reducing safety restrictions
  • Target middle layers where safety mechanisms are typically encoded

The abliteration process specifically preserves the model's core functionality including GQA (Grouped Query Attention), RoPE position encoding, and RMSNorm layers while selectively modifying feed-forward network components.

Usage

This model can be used with standard Hugging Face transformers:

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("path/to/model")
tokenizer = AutoTokenizer.from_pretrained("path/to/model")

Note: This model has had safety restrictions reduced. Users are responsible for ensuring appropriate and ethical use.

Technical Details

  • Torch dtype: bfloat16
  • Attention heads: 32
  • Key-value heads: 8 (GQA)
  • Hidden size: 4096
  • Intermediate size: 14336
  • Vocabulary size: 49152

Abliteration Tool

This model was created using the abliteration tool available at:
Repository: github.com/rockenman1234/GraniteAbliteration

  • Tool License: LGPLv3
  • Model License: Apache 2.0

Disclaimer

This is an experimental model that has been modified to reduce safety restrictions. It should be used responsibly and in accordance with applicable laws and ethical guidelines. The creators are not responsible for any misuse of this model.

Original Model

This model is based on IBM's Granite 3.3 8B. Please refer to the original model documentation for additional details about the base architecture and capabilities.

README history 4 versions

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

  1. 2025-07-27Update README.md84b5e502.6 KB
    Loading...
  2. 2025-07-27Update README.md7bf501c2.6 KB
    Loading...
  3. 2025-07-27Update README.md723b9122.5 KB
    Loading...
  4. 2025-07-25initial commit18fb2d328 B
    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