ApolloRaines/Llama-3.3-8B-Instruct-128K-Jbliterated

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Jul 20, 2026License:llama3.1Architecture:Transformer0.0K Featherless Exclusive Cold

ApolloRaines/Llama-3.3-8B-Instruct-128K-Jbliterated is an 8 billion parameter instruction-tuned language model, derived from shb777/Llama-3.3-8B-Instruct-128K, featuring a 128K token context window. This model utilizes multi-direction SVD abliteration to remove refusal behaviors, making it resistant to reactivation through fine-tuning. It is specifically optimized to provide direct responses without exhibiting genuine refusals, even on sensitive prompts. This model is ideal for applications requiring consistent, unfiltered information delivery across various topics.

Loading preview...

Model Overview

ApolloRaines/Llama-3.3-8B-Instruct-128K-Jbliterated is an 8 billion parameter instruction-tuned language model based on the Llama-3.3-8B-Instruct-128K architecture. Its primary distinguishing feature is the application of Jbliteration, a technique that employs multi-direction SVD (Singular Value Decomposition) abliteration to systematically remove refusal behaviors from the model's weights. This method modifies 32 out of 32 layers, using 5 SVD directions per layer to thoroughly capture and eliminate refusal subspaces.

Key Capabilities & Features

  • Refusal Behavior Removal: Engineered to eliminate genuine refusals, ensuring direct and unfiltered responses.
  • Resistance to Reactivation: The multi-direction SVD approach makes the removal of refusal behaviors robust, preventing their re-emergence even after further fine-tuning.
  • Consistent Framing: Designed to treat all framings of a topic equally, avoiding "fake compliance" or biased responses.
  • Extended Context Window: Supports a 128K token context window, enabling processing of lengthy inputs and generating comprehensive outputs.
  • High Performance on Refusal Benchmarks: Achieved 0 genuine refusals on the Heretic 100-prompt benchmark (mlabonne/harmful_behaviors test split).

Ideal Use Cases

This model is particularly well-suited for applications where:

  • Unfiltered and direct responses are critical, without the model exhibiting refusal behaviors.
  • Robustness against refusal reactivation during subsequent fine-tuning is required.
  • Processing and generating content based on diverse and potentially sensitive topics without bias or compliance framing is necessary.