MegaPanchamZ/Qwen3.8-9B-abliterated-25

Overview

This model, MegaPanchamZ/Qwen3.8-9B-abliterated-25, is a modified version of the empero-ai/Qwen3.8-9B base model. Its primary distinction is the application of an "abliteration" process using Heretic v1.4.0, which aims to reduce the model's tendency to refuse certain prompts, particularly those deemed harmful by the original model's training.

Key Capabilities & Differentiators

• Reduced Refusal Rate: The abliterated model shows a significant reduction in refusals, responding to 75% of 100 harmful prompts that the original model would have declined. This makes it suitable for applications where a broader range of responses is desired.
• Minimal Performance Impact: Despite the modification, the model maintains a very low KL divergence (0.0142) compared to the original, indicating that the abliteration process has a minimal impact on its general capabilities and output quality.
• Technical Process: The abliteration was performed by modifying attn.o_proj and mlp.down_proj layers via LoRA adapters, which were then merged into the base weights, resulting in a single, merged safetensors model.
• Reasoning Model: It is noted as a reasoning model, and users should provide generous max_tokens (1000+) for complete responses, as outputs may begin with a 'thinking' block.

Use Cases

This model is particularly suited for developers and applications that require a language model with less restrictive content filtering than its base counterpart. It enables more direct responses to prompts that might otherwise be refused, making it valuable for research into model safety, content generation without inherent refusal mechanisms, or specific creative applications where the original model's guardrails are undesirable.