← Back to Custom Model Training
2026-06-29Model TrainingQwenFine-tuningSLM

Qwen Franklin: Benjamin Franklin Fine-Tuned Model

A field note on accidentally turning fine-tuning, lore, and Qwen model variants into a Benjamin Franklin-flavored local-model rabbit hole.

Source context: This field note first appeared on Eric's Advisory Hour Substack. View the original post.

Screenshot of Qwen Franklin responding in Benjamin Franklin voice during a fine-tuned model experiment.
A Qwen Franklin run in character: the experiment starts with a tuned model that can speak back in Benjamin Franklin voice while still exposing practical reasoning.

I got into finetuning and lore creating custom Qwen models as an accident. It turns out this is a dangerous hobby. It takes everything about D&D character creation and worldbuilding, and then? It gives you a dopamine feedback loop that is impossible to break out of. I’ll have Ben add his comments as I write this article.

I’ll put those in these gray blocks.

I experimented with role-playing and printing in America and France. The useful result was the lightning-rod safety device.

The sheer vast size of the library, so to speak is staggering. One of the surprising things to me is just about many fine tunes and lora creations I’ve generated over such a short time period. I have a lot. I considered uploading them to huggingface, but it’s really tough to make sense of. Do you upload the best of all the trained models, just your preferred one for a given use-case? Does the world need qwen fine tuned ben franklin models?

Screenshot showing a growing local library of Qwen fine-tunes and LoRA model variants.
The local model library grew quickly: each LoRA or fine-tune became another candidate for a narrow job, another version to compare, and another artifact to explain.

I keep useful experiments and their best results. If you ask about a subject, I shall show your the evidence and reasoning behind the most useful answer.

SWE-Bench

The funny thing is, and this is not statistically meaningful by any stretch of the imagination, but qwen fine tuned as benjamin franklin seems to outperform qwen base on SWE-Bench. I haven’t re-run the benchmark many times, but it does appear that Franklin is 1-3% better on swe-bench than qwen without a lora+fine tuning and I can’t explain at all why that is.

Screenshot of a pseudo-SWE-Bench comparison graph for base Qwen and a Benjamin Franklin fine-tuned Qwen variant.
A pseudo-SWE-Bench result where the Franklin-tuned Qwen variant edges the base model enough to be interesting, even if not statistically decisive.

That is not to say it is great coding. You’re not coding a ThreeJS Wrigley Field with Qwen tuned Ben Franklin. However, it could be situationally useful in some narrow context that I’m just not aware of.

I’ve run into a lot of weird problems in these experiments. Here’s some of the better ones.

  1. The model started internal chain of thought switched to chinese. Franklin was a world traveler, but chinese is not likely what he reasoned internally with.
  2. Asking for javascript, you can get a short lecture on kites.
  3. Toolcalls are brittle-the various models struggle with these. Sometimes thinking happened in the thinking tags, and sometimes it was the response.

Thought for 1 second » I can help with that thought. The useful answer is Franklinized; the rest is boilerplum cutting. A simple HTML+JavaScript sketch would be better than a sermon on tool company names. Ask whether the user needs just the code or a practical explanation. Then close unless utility emerges.

Here’s a SWE-Bench setup. It’s not real SWE-Bench. That requires containers and more than I was willing to setup. This is pseudo-SWE-Bench. It uses the SWE-Bench problems just in a narrow fashion. The graph that emerges from the run is interesting and … also shows maybe I got a little carried away with custom models. :)

Screenshot of a SWE-Bench-style setup and model experiment graph for Qwen Franklin tests.
The experiment setup: SWE-Bench-style tasks, custom model variants, and enough runs to make the graph more useful than the anecdote alone.

So what can that qwn34b instruct with ben franklin lora do? Nothing earth shattering-but it can code, and it can do so at the lovely price of $0.00 per million tokens in or out.

Well, this is in one-shot. It’s pretty neat and has me thinking about taking all my threeJS projects and creating one kickass dataset.

Screenshot of a one-shot coding artifact generated by the Qwen Franklin model.
A one-shot code result from the Qwen Franklin model, useful less as a universal coder than as evidence that narrow, cheap model personalities can still produce working artifacts.

If you didn’t catch it, be sure to read about the Qwen Franklin Loras-which even includes a download link.

Related source

Summer into AI: Eric Submission #3: Qwen Franklin

Myles accelerated my timeline. I see you Myles and your “train a neural network in the browser” and “8-bit LLM” (links below). If I waited any longer I’m certain you’d stumble upon the same thought that I had.

Original source

This local copy preserves the post text, screenshots, and source link from Eric's Advisory Hour. Original Substack URL: https://advisoryhour.substack.com/p/qwen-franklin-benjamin-franklin-fine. You can also subscribe to Advisory Hour.