reaperdoesntknow/DistilQwen3-1.7B-uncensored

DistilQwen3-1.7B-uncensored Overview

This model, developed by Convergent Intelligence LLC: Research Division, is a 1.7 billion parameter language model from the DistilQwen3 Series. It is a product of a unique distillation chain based on Discrepancy Calculus (DISC), a measure-theoretic framework that decomposes the teacher's output distribution to quantify local structural mismatch. Unlike standard KL divergence, DISC focuses on local structural understanding.

Key Capabilities

• Proof-Weighted Knowledge Distillation: Utilizes a methodology where 55% cross-entropy with decaying proof weights (2.5x \u2192 1.5x) and 45% KL divergence at T=2.0 are applied. This amplifies loss on reasoning-critical tokens, forcing the student model to prioritize structural understanding.
• Optimized for Reasoning: Designed to allocate capacity to structural understanding rather than just surface-level pattern matching, making it effective for tasks requiring logical inference.
• BF16 Training: Part of the only BF16 collection in the Convergent Intelligence portfolio, trained on H100 with a 30B-parameter teacher, indicating a focus on high-fidelity compute.

Good for

• Instruction Following: As part of the DistilQwen collection, it benefits from distillation from a Qwen3-30B-A3B (Instruct) teacher, suggesting strong instruction-following capabilities.
• Structured Output & Legal Reasoning: The distillation process emphasizes structured output and legal reasoning, making it suitable for applications requiring precise and organized responses.
• STEM Derivation & Logical Inference: Its methodology, which focuses on structural decomposition and logical inference, makes it well-suited for STEM-related tasks and complex problem-solving.