saidutta69/Qwen2.5-1.5B-Instruct-heretic

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 2, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

saidutta69/Qwen2.5-1.5B-Instruct-heretic is a 1.5 billion parameter instruction-tuned causal language model, a decensored variant of Qwen/Qwen2.5-1.5B-Instruct. Developed by RACER IS OP using Heretic v1.2.0, it suppresses refusal behavior via targeted weight edits to attention and MLP down-projections, leaving base knowledge and instruction-following largely intact. This model is optimized for direct answers, making it suitable for local agents, roleplay, and research into alignment mechanics where over-refusal is a barrier.

Loading preview...

Overview

This model, saidutta69/Qwen2.5-1.5B-Instruct-heretic, is a 1.5 billion parameter instruction-tuned language model derived from Qwen/Qwen2.5-1.5B-Instruct. Its primary distinction is the suppression of refusal behavior through a process called "abliteration" (directional ablation) using Heretic v1.2.0. This method involves targeted weight edits to the attention output and MLP down-projections, rather than fine-tuning, ensuring the base model's knowledge and instruction-following capabilities remain largely preserved.

Key Capabilities

  • Decensored Responses: Significantly reduces refusal rates, dropping from 99/100 to 1/100 on adversarial prompts, allowing the model to answer directly.
  • Retained Core Functionality: Maintains the original Qwen2.5-1.5B-Instruct's knowledge and instruction-following due to the targeted nature of the edits (KL divergence of 0.1607).
  • Resource Efficient: At 1.5B parameters, it runs comfortably on CPUs or low-VRAM GPUs, outperforming the 0.5B variant in reasoning and coherence.
  • GGUF Quantizations: Provides a full suite of GGUF quantizations for various hardware configurations, including options for 6GB, 8GB, and 12GB GPUs, and CPU-only use.

Good For

  • Local Agents: Ideal for applications requiring a small, responsive model that provides direct answers.
  • Roleplay: Suitable for scenarios where models need to engage without excessive refusal.
  • Alignment Research: Useful for studying refusal mechanics and the impact of targeted weight edits on model behavior.
  • Use Cases Blocked by Over-Refusal: Any application where the default refusal behavior of RLHF-era models is a hindrance.

Finetunes
1 models