Distribution Fine Tuning: A post-training step to make models write better

Warning/Tips

• This is not a chatbot, and asking chat style questions won't work well
• It's designed to output long-form text, ask it to write an article/blog post/etc. about what you want to write.
• It will work best if you format your prompts like the examples.
• It was trained on web documents, so it will work best for blogs/articles/other similar writing.
• This is a small demo model trained on my home GPUs. Bigger/better models coming soon!
• Warning: to prevent blind copy-paste-post, this randomly replaces a word with a cute animal or fruit when you copy. Please read before sending it.

Acknowledge

DFT • writing demo

Distribution Fine Tuning: A post-training step to make models write better

Distribution Fine Tuning (DFT) is a new training algorithm that improves LLM outputs for writing. Read the Technical Report.

Tips: This model was trained on limited data on my home GPUs, so is much smaller than LLMs you might be used to. To get the best results, try to match the style of the examples as much as possible. If you are getting poor results, try using "Generate (good)" so the prompt is rewritten before generation. Make sure to experiment with the inputs to get the best possible output.

Be careful with the length parameter. If you set it too high or low compared to the amount of info in your outline, it can give poor results.

Write About

Write About

Prompt

Maximum 1024 tokens

Outline

Style

Writing style and tone

Use case

Intended context

Approximate output token length

↺

100 1000

Generate (good) Generate (fast) Stop Clear

DFT output

DFT

Examples

Example 1: Digital publishing and open access Example 2: Global water crisis advocacy Example 3: Teotihuacan cultural legacy

These examples set prompt, outline, style/use‑case, and target length.

What this is

This is a proof‑of‑concept writing demo (single prompt → single response). The model is 14B and trained with limited compute.

Generate (good) vs Generate (fast)

Generate (fast) calls the API directly with the prompt/outline as is. Generate (good) first rewrites the prompt/outline to match the style of the data it was trained on. If you match the style of the examples well, then you should be able to get good results with either.

Disclaimers

• It’s optimized for web‑text like news, blogs, and explanatory writing.
• Sometimes it may output irrelevant boilerplate (links/comments). If that happens, retry with a tighter prompt.
• Early demo, not a full product yet.

Contact

DM `@rosmine_b` on X (<https://x.com/rosmine%5Fb>) or email `hello@rosmine.ai`.

Get Updates

I'm currently working on building this into a "writing development environment": an app for people who want to use AI to make high quality writing, with generations, editing, and matching your style.

Put your e-mail here to get updates

Get updates