qwen3.5-2b-melvin-posts-v1-GGUF

Qwen3.5-2B Melvin Posts v1

A fine-tuned version of Qwen3.5-2B that turns topics and short briefs into casual X posts in Melvin’s writing style.

Trained with Unsloth Studio using LoRA on 324 curated examples derived from personal posts: 292 training examples and 32 validation examples. The dataset focuses primarily on local AI, coding agents, software development, and building projects.

This release includes reconstructed line breaks to encourage readable post formatting. Exported as a Q8_0 GGUF for local inference with LM Studio and other compatible llama.cpp-based applications.

Example

Usage

1. Turn off thinking

Open the right-hand Model Parameters panel.

Under Reasoning, switch Enable Thinking → Off.

This model is intended to write posts directly, without a reasoning response.

2. Set the system prompt

Find System Prompt and paste:

Write an X post in Melvin's voice from the supplied topic or brief. Use the supplied facts and opinions, keep the language natural and conversational, and return only the post. Do not invent experiences, claims, or links.

Keep this in the system prompt field—not in each chat message.

3.Generation settings

• Context length: 2048 — in model load settings.
• Enable Thinking: Off — under Reasoning.
• Temperature: 0.7 — under Settings.
• Limit Response Length: enable it and set to 256 tokens. Located below Temperature.

Expand Sampling and set:

• Top P: 0.8
• Top K: 20
• Min P: 0

Save these settings as a preset, then start a fresh chat.

Note: The response limit is measured in tokens, not characters. It does not guarantee the post fits X’s character limit.

These are starting settings, not requirements. 256 tokens is an output limit, not a 256-character limit.

Limitations

This is a small, specialized writing-style fine-tune—not a general-purpose assistant or factual reference. It may invent details, repeat training phrases, or produce inconsistent formatting. Review outputs before publishing. It is not guaranteed to stay within X’s character limit.