27 lines
1.0 KiB
Markdown
27 lines
1.0 KiB
Markdown
|
---
|
||
|
|
base_model: Qwen/Qwen2.5-3B
|
||
|
|
library_name: mlx
|
||
|
|
license: other
|
||
|
|
license_name: qwen-research
|
||
|
|
license_link: https://huggingface.co/Qwen/Qwen2.5-3B/blob/main/LICENSE
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
tags:
|
||
|
|
- mlx
|
||
|
|
- custom-kv-cache
|
||
|
|
- reasoning
|
||
|
|
---
|
||
|
|
|
||
|
|
# Nexus-Lumina-3B-v3
|
||
|
|
|
||
|
|
Nexus-Lumina-3B-v3 is a highly optimized, local reasoning engine fine-tuned by **GDI Nexus**.
|
||
|
|
|
||
|
|
To bypass instruction-tuning degradation (catastrophic forgetting), we fine-tuned the raw `Qwen2.5-3B` base model on 17,636 rigorously filtered reasoning chains from the OpenHermes dataset. It is engineered specifically for Apple Silicon via the MLX framework, utilizing our proprietary `mlx-ash-kv` (Asynchronous Self-Healing KV Cache) for maximum memory stability during extended contexts.
|
||
|
|
|
||
|
|
## Evaluation
|
||
|
|
Evaluated entirely locally on unified memory using the Hugging Face Leaderboard standard (25-shot):
|
||
|
|
* **Benchmark:** ARC-Challenge (25-shot)
|
||
|
|
* **Score:** 56.83% (Normalized Accuracy)
|
||
|
|
|
||
|
|
This model proves enterprise-grade 3B reasoning is achievable on local fanless hardware.
|