s3nh/fable-traces-abliterated

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

s3nh/fable-traces-abliterated is a decensored version of AliesTaha/fable-traces, which is a compact instruction-tuned language model built on Qwen/Qwen3-4B-Instruct-2507. This model is designed for short, conversational replies and runs efficiently on a single mid-range GPU. It has been modified using Heretic v1.4.0 to reduce refusals, maintaining the original model's refusal rate of 3/100 while introducing a KL divergence of 0.0011.

Loading preview...

Overview

s3nh/fable-traces-abliterated is a modified version of the AliesTaha/fable-traces model, which is based on the Qwen3-4B-Instruct-2507 architecture. This model has been specifically decensored using Heretic v1.4.0.

Key Characteristics

  • Decensored: Modified from its original version to alter refusal behavior, though it maintains the same refusal rate of 3/100 as the base model.
  • Compact Size: Built on a ~4 billion parameter base model, making it suitable for deployment on a single mid-range GPU.
  • Instruction-Tuned: Optimized for generating short, conversational replies.
  • Reproducible: The abliteration process is documented and reproducible, with specific parameters detailed in the reproduce directory.

Performance & Details

While the original model had a KL divergence of 0 (by definition), this abliterated version introduces a KL divergence of 0.0011. It uses the ChatML prompt format and inherits the context length of its base model. The model is distributed in bfloat16 precision using safetensors.

Usage

This model can be used with the transformers library for inference or served efficiently using vLLM.

Finetunes
1 models