Overview

This model, llmfan46/GLM-4-32B-0414-uncensored-heretic-v1, is a decensored variant of the original zai-org/GLM-4-32B-0414 model. Developed by llmfan46 using the Heretic v1.2.0 tool and the Arbitrary-Rank Ablation (ARA) method, its primary distinction is a 90% reduction in refusals (10/100 compared to 100/100 for the original) while maintaining model quality with a KL divergence of 0.0200.

Key Capabilities

• Decensored Output: Significantly fewer content restrictions and refusals.
• Function Calling: Supports external tool calls in JSON format, demonstrated with examples for HuggingFace Transformers.
• Code Generation: Achieves strong performance in engineering code and artifact generation, including animation and web design.
• Search-Based Q&A: Excels in generating detailed analytical reports based on provided search results.
• Reasoning: The base GLM-4-32B-0414 model, on which this is built, was pre-trained on 15T of high-quality data, including substantial reasoning-type synthetic data.

Performance Highlights

• Refusals: 10/100 (compared to 100/100 for the original).
• KL Divergence: 0.0200 (indicating high preservation of original model capabilities).
• Benchmarks: Outperforms Qwen2.5-Max and DeepSeek-R1 on IFEval, BFCL-v3, TAU-Bench, SimpleQA, and HotpotQA. Achieves 33.8 on SWE-bench Verified with Moatless.

Good For

• Use cases requiring reduced content restrictions and more direct responses.
• Code generation and function calling applications.
• Complex Q&A and report generation tasks leveraging search results.
• Scenarios where preserving the original model's quality while enhancing uncensored behavior is critical.