reaperdoesntknow/Gemma-3-270m-Opus-Distil
reaperdoesntknow/Gemma-3-270m-Opus-Distil is a 0.3 billion parameter Gemma 3 family model fine-tuned by Convergent Intelligence LLC. This experimental checkpoint is specifically adapted for reasoning-style English text generation using a sparse fine-tuning setup and the custom CIxOpt optimizer framework. It is intended for research into efficient model adaptation and small-model reasoning experiments, offering a compact solution for targeted text generation tasks.
Loading preview...
Model Overview
This model, reaperdoesntknow/Gemma-3-270m-Opus-Distil, is an experimental 270 million parameter Gemma 3 derivative developed by Convergent Intelligence LLC. It has been fine-tuned from google/gemma-3-270m using a unique sparse fine-tuning setup and the custom CIxOpt optimizer framework.
Key Differentiators & Capabilities
- Reasoning-Style Adaptation: Specifically trained on
angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7kto shape its output towards reasoning-style text generation. - Sparse Fine-Tuning: Employs a strategy of selective parameter participation, aiming to adapt specific reasoning and response-shaping surfaces while preserving the compact pretrained backbone and avoiding full-model disturbance.
- CIxOpt Optimizer: Utilizes a custom, heterogeneous optimizer designed for architecture-aware routing, supporting various update styles (AdamW, Lion, AdaMax) and parameter-name-aware routing for efficient adaptation.
- Experimental Focus: Primarily intended for research into compact Gemma fine-tuning, optimizer experiments, and efficient adaptation strategies rather than a general-purpose assistant.
Intended Use Cases
- Research on compact Gemma fine-tuning and optimizer behavior.
- Experiments with small-model reasoning-style generation.
- Local text generation and instruction-following studies.
- Prototyping and testing small agent backbones.
- Educational analysis of model behavior and efficient adaptation.
Limitations
As an experimental checkpoint, it may exhibit limitations such as hallucination, sensitivity to prompt format, and potential for repetition. It has not been fully evaluated for factuality, safety, or complex tasks, and its small size inherently limits world knowledge and reasoning depth.