πŸš€ Singularity-Max (Qwen3.5-4B) - Stage 3.1 Conservative Edition

πŸ“Œ Overview

This model is the ultimate realization of the Singularity-aware Adaptive Surgery. Moving beyond naive uniform pruning (which inherently degrades mathematical reasoning), this model employs a strictly conservative, layer-wise adaptive pruning strategy based on the physical horizon of the trace-class singular spectrum.

Key Breakthroughs:

  1. q_proj Sanctuary & Structural Core Protection: All MLPs, Norms, Embeddings, and q_proj layers are 100% preserved to maintain instruction following and attention steering.
  2. Stair-step Pruning Logic: Pruning ratios dynamically shift guided strictly by physical noise boundaries, guaranteeing zero intrusion into the dense parameter core.
  3. Risk Fusion Validation: Targets were surgically selected out of a vast candidate pool using a multi-dimensional risk fusion equation.

This guarantees zero degradation in mathematical reasoning (perfectly outputs 5 for 2x + 5 = 15) and completely prevents <think> tag leakage, while successfully purging non-zero flux condition (NZFC) noise tails.

βš–οΈ License

Licensed under CC BY-NC 4.0.

Downloads last month
184
Safetensors
Model size
4B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for SingularityPrinciple/Qwen3.5-4B-Singularity-Max

Finetuned
Qwen/Qwen3.5-4B
Finetuned
(40)
this model
Quantizations
2 models