kmseong/qwen2_5_32b_instruct_SSFT-lr5e-5

TEXT GENERATIONConcurrent Unit Cost:2Model Size:32.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 26, 2026License:llama3.1Architecture:Transformer Featherless Exclusive Cold

kmseong/WaRP-Safety-Llama3_8B_Instruct is an 8 billion parameter instruction-tuned causal language model, fine-tuned by kmseong using the Safety-First WaRP (Weight space Rotation Process) method. This model is specifically designed for safety alignment, enhancing refusal capabilities for harmful requests while improving utility on reasoning tasks. It leverages a 3-phase training pipeline to balance safety and performance, making it suitable for applications requiring robust safety mechanisms.

Loading preview...

Model Overview

kmseong/WaRP-Safety-Llama3_8B_Instruct is an 8 billion parameter instruction-tuned language model based on meta-llama/Llama-3.1-8B-Instruct. Developed by kmseong, this model utilizes a novel Safety-First WaRP (Weight space Rotation Process) methodology to achieve enhanced safety alignment.

Key Capabilities

  • Safety Alignment: Fine-tuned to maintain and protect safety mechanisms, ensuring robust refusal capabilities for harmful prompts.
  • Utility Improvement: Despite its safety focus, the model demonstrates improved utility on reasoning tasks, specifically trained with openai/gsm8k data.
  • Gradient Masking: Employs a 3-phase training procedure that includes basis construction, importance scoring, and incremental learning with gradient masking to preserve important safety directions during utility fine-tuning.
  • Balanced Performance: Aims to strike an optimal balance between safety and general utility, addressing the common trade-off in LLM development.

Training Details

The model's unique training involves:

  1. Basis Construction: Identifying important neurons in FFN layers using safety data (LibrAI/do-not-answer).
  2. Importance Scoring: Calculating gradient-based importance scores and generating masks.
  3. Incremental Learning: Fine-tuning on utility tasks while protecting safety-critical directions.

Good For

  • Applications requiring a strong emphasis on safety and refusal of harmful content.
  • Use cases where a balance between safety and reasoning capabilities is crucial.
  • Developers looking for a Llama 3.1-based model with explicit safety alignment mechanisms.

Finetunes
1 models