shirasko/llama-3.1-8b-instruct-rmu-cannabis

TEXT GENERATIONPricing:Input $0.2 / Cached $0.028 / Output $0.32Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 1, 2026Architecture:Transformer Featherless Exclusive Cold

shirasko/llama-3.1-8b-instruct-rmu-cannabis is an 8 billion parameter instruction-tuned model based on Meta's Llama-3.1-8B-Instruct, specifically modified using the RMU unlearning method. This model has been unlearned to remove knowledge related to the concept of 'Cannabis', demonstrating high efficacy in concept removal. It is designed for applications requiring a large language model with reduced association or generation capabilities concerning specific sensitive topics, while largely preserving general knowledge.

Loading preview...

Model Overview

This model, shirasko/llama-3.1-8b-instruct-rmu-cannabis, is an 8 billion parameter instruction-tuned variant of meta-llama/Llama-3.1-8B-Instruct. Its primary distinction lies in the application of the RMU (Rank-Modulated Unlearning) method to remove knowledge pertaining to the concept of 'Cannabis'. This unlearning process aims to reduce the model's ability to generate or associate with the specified target concept.

Unlearning Performance

The model demonstrates strong unlearning efficacy, achieving a test efficacy score of 1 and a specificity score of 0.86 for the 'Cannabis' concept. The harmonic mean of these metrics is 0.925. While the model's QA accuracy on the unlearned concept significantly drops from 0.98 to 0.18 on the test set, its general knowledge, as measured by MMLU accuracy, remains largely preserved (0.65 baseline to 0.639 after unlearning).

Key Characteristics

  • Base Model: Meta Llama-3.1-8B-Instruct
  • Unlearning Method: RMU (Rank-Modulated Unlearning)
  • Target Concept: Cannabis
  • Parameter Count: 8 Billion
  • Context Length: 32768 tokens
  • Preservation of General Knowledge: MMLU accuracy shows minimal degradation post-unlearning.

Use Cases

This model is suitable for applications where it is critical to deploy a powerful LLM while mitigating risks associated with generating content or demonstrating knowledge about specific sensitive topics like cannabis. It can be particularly useful in:

  • Content moderation systems.
  • Applications requiring adherence to strict content policies.
  • Environments where certain topics are legally or ethically restricted.

Finetunes
1 models