reaperdoesntknow/DistilQwen3-1.7B-uncensored

Model Overview

This model, reaperdoesntknow/DistilQwen3-1.7B-uncensored, is a 1.7 billion parameter language model developed by Convergent Intelligence LLC: Research Division. It is part of the DistilQwen Collection, a series of models built using a unique proof-weighted knowledge distillation methodology based on Discrepancy Calculus (DISC).

Key Capabilities & Methodology

• Distillation from a 30B-parameter Teacher: This model is distilled from a powerful Qwen3-30B-A3B teacher, leveraging premium H100 hardware and BF16 training for enhanced performance.
• Discrepancy Calculus (DISC): Utilizes a measure-theoretic framework to decompose the teacher's output distribution, quantifying local structural mismatch that standard KL divergence misses. This approach aims to improve the student model's understanding of underlying structure.
• Proof-Weighted Knowledge Distillation: Employs a unique distillation process (55% cross-entropy with decaying proof weights, 45% KL divergence at T=2.0). The proof weight specifically amplifies loss on reasoning-critical tokens, compelling the student to prioritize structural understanding over superficial pattern recognition.
• Focus on Reasoning: Designed to excel in tasks requiring deep reasoning, logical inference, and structural decomposition, as opposed to mere surface-level pattern matching.

DistilQwen Collection Context

This model is part of a collection that includes variants optimized for different strengths, such as instruction following, structured output, legal reasoning, extended deliberation, higher-entropy distributions, proof derivation, STEM derivation, and logical inference. The entire DistilQwen series is notable for being the only BF16 collection within the Convergent Intelligence portfolio, indicating a higher compute investment compared to their broader catalog.