reaperdoesntknow/DistilQwen3-1.7B-uncensored
reaperdoesntknow/DistilQwen3-1.7B-uncensored is a 1.7 billion parameter uncensored language model from the DistilQwen3 series by Convergent Intelligence LLC: Research Division. This model is part of a proof-weighted distillation chain built on Discrepancy Calculus, designed to amplify reasoning-critical tokens. It excels in instruction following, structured output, and legal reasoning, leveraging a 30B-parameter teacher model trained on H100 at BF16.
Loading preview...
Model Overview
reaperdoesntknow/DistilQwen3-1.7B-uncensored is a 1.7 billion parameter model from the DistilQwen3 series by Convergent Intelligence LLC: Research Division. It is part of a proof-weighted distillation chain, utilizing Discrepancy Calculus to enhance structural understanding over surface-level pattern matching. This model was distilled from a 30B-parameter teacher model using premium hardware (H100 at BF16).
Key Capabilities
- Instruction Following: Optimized for precise adherence to given instructions.
- Structured Output: Capable of generating well-organized and structured responses.
- Legal Reasoning: Shows proficiency in tasks requiring legal reasoning.
- Proof-Weighted Distillation: Employs a unique methodology that amplifies loss on reasoning-critical tokens, fostering deeper understanding.
Methodology
The DistilQwen series, including this model, is the only BF16 collection in the Convergent Intelligence portfolio, trained on H100 GPUs. It uses a proof-weighted knowledge distillation approach, combining 55% cross-entropy with decaying proof weights (2.5x \u2192 1.5x) and 45% KL divergence at T=2.0. This method forces the student model to prioritize structural understanding, as detailed in "Structure Over Scale" (DOI: 10.57967/hf/8165).
Top 3 parameter combinations used by Featherless users for this model. Click a tab to see each config.