reaperdoesntknow/DistilQwen3-1.7B-uncensored

Model Overview

reaperdoesntknow/DistilQwen3-1.7B-uncensored is a 1.7 billion parameter model developed by Convergent Intelligence LLC: Research Division. It is a key component of the DistilQwen Collection, a proof-weighted distillation series. Unlike other models in the broader Convergent Intelligence catalog, the DistilQwen series was trained on premium H100 hardware at BF16 precision, distilling knowledge from a 30B-parameter teacher model.

Mathematical Foundations

This model's distillation chain is built upon Discrepancy Calculus (DISC), a measure-theoretic framework that decomposes a teacher's output distribution. The discrepancy operator quantifies local structural mismatch, which is crucial for robust knowledge transfer. The full theory is detailed in "On the Formal Analysis of Discrepancy Calculus" (Colca, 2026) and the methodology in Structure Over Scale (DOI: 10.57967/hf/8165).

Distillation Methodology

All DistilQwen models utilize a proof-weighted knowledge distillation approach:

• 55% cross-entropy with decaying proof weights (2.5x to 1.5x).
• 45% KL divergence at T=2.0.

This method amplifies loss on reasoning-critical tokens, compelling the student model to prioritize structural understanding and logical inference rather than mere surface-level pattern recognition. This makes the model particularly suited for tasks requiring deeper comprehension and reasoning.