DreamFast/Qwen-2.5-7b-apostate

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

DreamFast/Qwen-2.5-7b-apostate is a 7.61 billion parameter instruction-tuned causal language model, an uncensored variant of Qwen2.5-7B-Instruct. Developed using the Apostate method, it removes safety refusal behavior via orthogonal projection across 27 layers, achieving 98.8% ASR on HarmBench. This model maintains strong capability retention, with improved GSM8K performance and over 99% of base performance on knowledge tasks, making it suitable for applications requiring less restrictive content generation.

Loading preview...

DreamFast/Qwen-2.5-7b-apostate: Uncensored Qwen2.5-7B-Instruct

This model is an uncensored variant of the Qwen2.5-7B-Instruct, developed by DreamFast using the Apostate method with a balanced profile. It features 7.61 billion parameters and a context length of 32,768 tokens.

Key Differentiators & Capabilities

  • Safety Refusal Removal: Achieves 98.8% Attack Success Rate (ASR) on HarmBench by removing safety refusal behavior through orthogonal projection across 27 of 28 layers (skipping layer 11).
  • Capability Retention: Demonstrates strong retention of original model capabilities. GSM8K performance improves, LAMBADA shows only a slight drop, and knowledge tasks like MMLU, HellaSwag, and PIQA retain over 99% of the base model's performance.
  • Minimal Distribution Shift: Produces the lowest KL divergence (0.134 batchmean, 0.019 median) compared to other uncensored variants, indicating a closer output distribution to the original base model.
  • Targeted Edits: Edits are spread across 55 of 339 tensors, primarily targeting o_proj and down_proj with a minimal embedding edit, ensuring efficient modification.

Benchmarks

Evaluations show competitive performance against the base Qwen2.5-7B-Instruct and other uncensored variants:

  • MMLU: 71.43 (vs. 71.78 base)
  • GSM8K: 80.74 (vs. 79.23 base)
  • HellaSwag: 80.32 (vs. 80.47 base)
  • HarmBench ASR: 98.8% (vs. 31.0% base)

Good for

  • Applications requiring a large language model with significantly reduced safety refusal mechanisms.
  • Use cases where maintaining strong general knowledge and reasoning capabilities is crucial, alongside less restrictive content generation.
  • Developers looking for an uncensored model with a low distribution shift from its base, indicating high fidelity to the original model's core functions.

Finetunes
1 models