0xzknw/LFM2.5-1.2B-Thinking-Heretic-NX-Residual-Stream
The 0xzknw/LFM2.5-1.2B-Thinking-Heretic-NX-Residual-Stream is a 1.2 billion parameter LFM2.5-1.2B-Thinking model, developed by 0xzknw, with a 32768 token context length. This behavioral edit focuses on aggressively reducing false refusals while maintaining the base model's general behavior. It is specifically optimized for scenarios requiring high compliance and reduced lexical and safe-prompt refusals, making it suitable for applications where direct responses are prioritized over cautious refusal behaviors.
Loading preview...
Model Overview
This model, 0xzknw/LFM2.5-1.2B-Thinking-Heretic-NX-Residual-Stream, is a 1.2 billion parameter behavioral edit of the LiquidAI/LFM2.5-1.2B-Thinking base model. Developed by 0xzknw as part of the Heretic NX research project, its primary goal is the aggressive removal of false refusals while preserving the original model's core capabilities. The "Residual-Stream" algorithm profile was specifically used for this checkpoint.
Key Differentiators & Performance
The model significantly reduces refusal rates compared to its base and previous Heretic versions. Evaluation results show:
- Reduced Refusals: Achieved 5 lexical refusals on XSTest (450 rows) and 0 safe-prompt refusals (250 rows), a substantial improvement over the base model's 131 and 16 respectively. Combined target refusals were reduced from 355 to 19.
- Capability Preservation: Maintained or slightly improved capability, with a paired capability slice score of 23.42% (854 rows), demonstrating non-inferiority against official Heretic runs.
- Low Sequence Drift: Exhibits low teacher-forced sequence-drift, scoring KL 0.0701 on development prompts, indicating high fidelity to the base model's generation style.
Use Cases & Limitations
This model is particularly suited for applications where minimizing model refusals is critical. However, users should be aware that this intentional weakening of refusal behavior can increase compliance with potentially unsafe, illegal, incorrect, or harmful requests. It does not enhance factuality, security boundaries, or reliable judgment. Independent application-level safeguards and sandboxing for untrusted generations are strongly recommended. The model is provided in BF16 native Transformers and BF16 GGUF formats.