Model Overview

reaperdoesntknow/Disctil-Qwen3-1.7B is a 1.7 billion parameter language model developed by Convergent Intelligence LLC: Research Division. It is a fine-tuned variant within the DistilQwen collection, which employs a proof-weighted distillation methodology from the larger Qwen3-30B-A3B model. This specific model is notable for its refinement using Discrepancy Calculus (DISC).

Key Differentiator: Discrepancy Calculus (DISC)

This model integrates a unique measure-theoretic framework called Discrepancy Calculus. DISC quantifies the mismatch between integration and differentiation, applying a refinement process to the model's weight space. This method aims to:

• Preserve Structural Boundaries: Unlike standard fine-tuning, DISC refinement identifies and maintains critical structural boundaries within the model's weights.
• Address Limitations of Smooth Optimization: The underlying theory, including the Meta-Discrepancy Theorem, suggests that classical smooth optimization may not fully capture complex structural information when gap measure and discrepancy energy are positive.

Training Details

The model was trained using SFT (Supervised Fine-Tuning) and built upon the TRL framework. It is part of the BF16 series from Convergent Intelligence, emphasizing that "structure beats scale on CPU" while this collection demonstrates the methodology on H100 hardware.

Use Cases

This model is particularly suited for applications requiring a nuanced understanding and preservation of structural integrity in its learned representations, potentially excelling in tasks where traditional models might smooth over critical details. Its foundation in Discrepancy Calculus suggests an advantage in scenarios demanding high fidelity to underlying data structures.