Train Your Own Decision Model Locally with Unsloth
Melvin Vivas · X post · 2026-10-08 · Open on X
Topics: Fine-tuning & Model Customization · Level: intermediate
Summary
Melvin shares news that you can now train your own 'decision model' on a base LLM such as Qwen3.5 or Gemma 4 using Unsloth's open-source repo. The quoted post says fine-tuning Qwen3.5 0.8B raised its aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks, using only 4GB of VRAM.
Key points
- Decision models can be fine-tuned from small open base models like Qwen3.5 or Gemma 4.
- The Qwen3.5 0.8B result: aggregate accuracy rose from 20.7% to 74.3% across 3 decision benchmarks.
- Training ran locally on just 4GB of VRAM.
- The tooling is Unsloth's open-source repo.
Resources mentioned
- Unsloth · tool · unsloth.ai · free
Open-source library for fast, memory-efficient fine-tuning of open LLMs (LoRA/QLoRA) on a single GPU or in Colab.
Also in: 100-Iteration Autonomous Coding Loop with Local Qwen on an RTX 3090 (Melvin Vivas on X · notes), Prompting an Autonomous Agent Loop: Be Specific, Show Visible Progress (Melvin Vivas on X · notes), Running an Overnight 100-Iteration Agent Loop with a Local Qwen Model (Melvin Vivas on X · notes), Agent tasks: Unsloth fine-tuning, ComfyUI face swap, script writing (Melvin Vivas on X · notes) and 36 more - Qwen 3.5 · tool · qwen.ai · free
Alibaba Qwen open model family with small on-device variants and toggleable reasoning.
Also in: Workshop AI: Building Apps with Cloud and Local Agents (GLM 5, Qwen 3.5) (Melvin Vivas on X · notes), Qwen3.5 0.8B Does Real-Time Local Video Captioning (Melvin Vivas on X · notes), Qwen3.5 0.8B Real-Time Video Captioning on Mac Studio (Melvin Vivas on X · notes), Qwen 3.5 Is the Local Model That Holds Up in Claude Code (Melvin Vivas on X · notes) and 8 more - Gemma 4 · tool · ai.google.dev · free
Google's family of open-weight models in several sizes, built to run on devices and offline, with multimodal and agentic abilities, and open to fine-tuning.
Also in: Meetily: Open-Source Local AI Meeting Notes App (Melvin Vivas on X · notes), New Gemma Developer Documentation Site (Melvin Vivas on X · notes), A100 80GB GPUs in Google Colab via Google AI Pro plans (Melvin Vivas on X · notes), A day of work done from a phone with Grok Bot agents (Melvin Vivas on X · notes) and 32 more
Try this
- Try fine-tuning a small base model (Qwen3.5 0.8B or Gemma 4) into a decision model with Unsloth.
- Fine-tune Qwen3.5 0.8B locally on a 4GB GPU into a decision model and compare its benchmark accuracy before and after.
More in Fine-tuning & Model Customization
- Fine-tune and Run Local Models for Free on Google Colab's T4 GPU
- Free Colab notebooks for fine-tuning and running local models
- Agent tasks: Unsloth fine-tuning, ComfyUI face swap, script writing
- Using an ML agent to train an open-source TTS model on your voice
- Run Local Models on a Free GPU with Google Colab (T4)
- Free Colab Notebooks for Local Model Fine-Tuning and Inference