LeadFootThrottleCock/Qwen2.5-7B-Instruct-heretic

TEXT GENERATIONConcurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Mar 16, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

LeadFootThrottleCock/Qwen2.5-7B-Instruct-heretic is a 7.6 billion parameter instruction-tuned causal language model, based on Qwen's Qwen2.5-7B-Instruct, that has been abliterated (decensored) using the Heretic tool. This model retains the original capabilities of the base Qwen2.5-7B-Instruct while significantly reducing refusal behaviors, making it suitable for creative writing with mature themes and balanced discussions of controversial topics. It offers a 32K context length and is provided in various GGUF quantizations for efficient deployment on consumer hardware, including AMD ROCm-compatible systems.

Loading preview...

Model Overview

LeadFootThrottleCock/Qwen2.5-7B-Instruct-heretic is a 7.6 billion parameter instruction-tuned model derived from Qwen's Qwen2.5-7B-Instruct. Its primary distinction is the "abliteration" process, which decensors the base model using the Heretic v1.2.0 tool. This process aims to remove refusal behaviors while preserving the model's core capabilities, as indicated by a low KL Divergence of 0.0820.

Key Capabilities & Features

  • Decensored Output: Designed to engage in creative writing with mature themes, factual discussions without hedging, and balanced handling of controversial topics without moralizing.
  • High Fidelity: Achieves minimal capability degradation from the base Qwen2.5-7B-Instruct model.
  • GGUF Quantizations: Provided in various GGUF formats (BF16, Q8_0, Q6_K, Q5_K_M, Q4_K_M) for flexible deployment across different hardware configurations.
  • ROCm Compatibility: Abliteration was performed with specific patches to Heretic to ensure correct operation on AMD RDNA3 GPUs, addressing issues like pad token handling and SDPA backend problems.
  • ChatML Template: Utilizes the standard Qwen2.5 ChatML format for instruction following.

Recommended Use Cases

  • Creative Writing: Ideal for generating content with mature or unrestricted themes.
  • Unfiltered Information Retrieval: Suitable for tasks requiring direct answers on sensitive or controversial subjects without built-in moralizing.
  • General Instruction Following: Retains full capability for standard coding tasks and factual knowledge queries.
  • Local Deployment: Optimized GGUF quantizations make it suitable for running on consumer-grade hardware, including AMD GPUs.

Finetunes
2 models