richardyoung/fable-qwen2.5-3b-agentic-merged-heretic
The richardyoung/fable-qwen2.5-3b-agentic-merged-heretic is a 3.1 billion parameter Qwen2.5-based causal language model, derived from devxyasir/fable-qwen2.5-3b-agentic-merged and decensored using Heretic v1.4.0. This model is specifically modified to reduce refusals, offering a less restrictive response generation compared to its original counterpart. It is suitable for applications requiring more open-ended or less filtered AI interactions.
Loading preview...
Model Overview
This model, fable-qwen2.5-3b-agentic-merged-heretic, is a 3.1 billion parameter Qwen2.5-based causal language model. It is a decensored version of the devxyasir/fable-qwen2.5-3b-agentic-merged model, created using the Heretic v1.4.0 tool. The original model was developed by devxyasir and fine-tuned from unsloth/qwen2.5-3b-instruct-unsloth-bnb-4bit, leveraging Unsloth for faster training.
Key Differentiator
The primary distinction of this 'Heretic' version is its significantly reduced refusal rate. While the original model exhibited 96 refusals out of 100 test cases, this decensored variant shows only 3 refusals out of 100. This makes it suitable for use cases where a more direct and less restrictive response generation is desired.
Reproducibility
The model's creation process is reproducible, with detailed instructions available in the reproduce directory.
Performance Metrics
- KL divergence: 0.0503 (compared to 0 for the original model by definition)
- Refusals: 3/100 (compared to 96/100 for the original model)
Training Details
The base model was fine-tuned using Unsloth and Huggingface's TRL library, enabling 2x faster training.