Field report · four training runs · June 2026
Simplified Technical English didn't do it. Zinsser didn't do it. Google's style guide didn't do it. So I spent three weeks and $113 training a model to sound like the engineers I like reading. A one-line prompt to Claude beat it.
Forty dollars of rented H100 time, and my model finally spoke. It opened with a byline, for an article nobody had written, then repeated its own title and dropped a stray HTML comment into the middle of a paragraph.
That was run one of four.
The plan had a clean shape. A frontier model knows things, so let it carry the facts. A small model, tuned on good human technical writing, would carry the prose. Knowledge from the big one, voice from the little one.
I ran the idea past Claude before I spent a dollar. It told me, politely, that I would lose: a well-prompted frontier model would out-write a fine-tuned small one, because it is already the better writer and prompting costs nothing next to training. It was a good argument. I built the thing anyway, because I did not want to be told.
It lost. And I would spend the $113 again, which is what this post is about.
The rulebooks I tried first
The technical prose these models produce is slop. Padded, hedged, structurally identical from one answer to the next, every point wearing a bold label. Once you have noticed the shape you cannot stop noticing it, and I was fighting it in every draft.
So I did not start by renting a GPU. I started with the rulebooks, because people have been codifying good technical writing for fifty years and I assumed somebody had already written the answer down.
ASD-STE100, Simplified Technical English, is the strictest: roughly nine hundred approved words, one meaning per word, short sentences. It exists so a technician reading a maintenance procedure in their third language, at two in the morning, on an aircraft, cannot misread it. At that job it is superb. It also strips the voice out on purpose, which is the entire point of it and exactly my problem.
Google's developer documentation style guide makes a large team's output consistent, which is hard and worth doing. Consistent and alive are different things.
Zinsser comes closest. But when you codify On Writing Well into something a model can follow, the half that survives is the subtractive half. Cut the adverb. Cut the qualifier. Those are rules. "Build a paragraph that carries an argument" is a craft, and it does not fit in a rubric.
Hold on to that. I am about to make the same mistake, at much greater expense.
Each of those helped, and none got me where I wanted. The output came back shorter, tighter, more correct and no more alive, because "don't sound like a machine" is a far weaker instruction than "sound like this person." You cannot write down what a good voice is. You can point at a few hundred examples of one. So point.
Eight stages, and how each one lies to you
Tuning a twelve-billion-parameter model is a sequence, and almost nothing in the sequence fails loudly. The model rarely crashes. It hands you something slightly wrong and lets you discover it much later, usually after you have paid for the run that produced it. Every failure below happened to me.
The eight stages, and where each one lies
Left: what the stage does. Right: how it goes wrong while reporting success.
Your model is a mirror of your corpus
Run one finished cleanly and produced garbage. It wrote author bylines, repeated its own title, left HTML comments sitting in the prose. I had scraped a few hundred sources and then taught the model, carefully and at some expense, to imitate the scrape.
The audit was humbling. My data preparation code was prepending a markdown byline to 89% of the training documents. Nearly a fifth of my instruction examples ran longer than the context window, so the trainer cut their tails off and took the stop token with them. That is why the model never knew when to stop talking.
The subtler failure showed up after I had fixed all of that. The corpus passed every cleanliness check I could write and was still wrong, because four-fifths of it by word count was dry specification text. I had assembled it for coverage when the model's entire job was register. Rewriting the weighting took readable human voice from 16.5% of the corpus to 59.7%, and made it smaller doing it: 11,546 packed blocks down to 7,034.
Then someone asked me a plain question mid-run: will it stop answering with links? I went and looked instead of guessing, and 84% of my instruction answers contained inline markdown links, six apiece at the median. I killed the run there, two dollars in, stripped the URLs and started over. Letting it finish would have cost $9 and produced something I would have thrown away. One plain question was worth more than every automated check I had written.
The run that went perfectly and produced a wreck
The second attempt trained to completion, published its result and reported success. Final evaluation loss of 7.37, worse than where an untrained model starts. Training dropped the way it should, bottomed out around 1.48 by step 30, then blew up and stayed up for the remaining seven hundred steps. A learning rate one notch too high. Nothing errored. It spent the money and handed me the wreck with a clean exit code.
Three separate safety nets should have caught this. All three were looking somewhere else.
Where the guards were looking
Run v2b, lr 2e-4. Loss values are from the run's history; guard positions are from the config.
A smoke test that ends before your failure mode starts only reassures you. So I stopped and rebuilt the harness before spending another dollar: a kill switch inside the training loop, milestones with the billing switched off between them, and uploads verified by re-reading the destination rather than trusting an exit code.
What it cost, run by run
Three of the four runs produced nothing I kept. That ratio is roughly what learning something looks like.
Four runs, one usable model
Rented H100 time at $3.29/hour, June 2026.
The test that actually mattered
The third run trained clean and measurably cut the machine-writing tics: 0.91 on my LLM-speak metric against a stock Gemma's 1.33, winning thirteen of twenty held-out briefs. I had built a voice tool rather than a smarter model, which is what I set out to build.
Then I gave the same notes and the same plain instruction to three writers: Claude Sonnet, a stock Gemma 4 26B, and mine. Then I read all three.
Claude Sonnet
Frontier · plain prompt · 972 words
"When you stuff a complex task into a single prompt, you're asking the model to do too many things at once. It will drop instructions, mix up concerns that should stay separate, and produce output that's difficult to debug."
Every word inside a paragraph. Not one bullet in the whole chapter.
Gemma 4 26B
Stock · plain prompt · 874 words
"…you gain three massive advantages:
* Reliability: Each call is simpler…
* Observability: You can inspect and test…
* Granular Control: You aren't locked in."
Opens with good paragraphs, then reaches for a bold label every other line. More than half its words live inside bullets.
My fine-tune
Tuned for voice · 422 words
"Your LLM prompt is too big and unwieldy for the task."
"Your prompt has a lot of conflicting instructions."
"Your prompt's output is hard to parse."
No filler, no hedging, and no paragraphs either. I asked for a chapter and got a list of sentences standing on their own.
Average length of a paragraph, in words
Measured on the three chapter files. Headings excluded.
So the model I trained specifically for voice lost on voice, to a model I simply asked nicely.
Why it lost
I had taught it the easy half of writing well. It dropped the hedges, the throat-clearing, the "it's important to note." Those are words, and words are cheap to count, so words are what my scoring rewarded.
Structure is the hard half, and structure is the deeper tell. Bulleting everything is what makes machine writing feel like machine writing, more than any individual word does, and my corpus was still full of documents that bullet everything. It taught the structure right back in while I was congratulating myself about vocabulary. I optimized the symptom I could count and missed the one I could not. Which is the same mistake the rulebooks make, run again at $113 a go. Subtraction fits inside a rule. Construction needs judgment.
What I owe the people who do this for a living
Every mistake here is an ordinary, known thing to someone who trains models for a living. A practitioner would have started the learning rate lower and clipped the gradients without thinking about it. A clean dataset can still be the wrong dataset, which is the first thing anybody who has built one will tell you. I paid $113 to learn a page of things a good team already knows in their hands.
And every choice in the recipe is a hypothesis. I made reasonable guesses and ran four configurations. The people who are good at this run hundreds, changing one variable at a time, and most of those runs teach them nothing except that a door is closed. Patient, largely unrewarded, mostly negative-result work. That is the actual job.
Almost none of what I used did I pay for. The loaders, the trainers, the quantiser, the hub the model sits on: built and given away by people who did not have to. So were the essays my corpus is made of, a few hundred engineers who wrote carefully in public for no particular reward. Whatever voice my model has, it is theirs. I borrowed it.
What $113 bought me
The eight rules I would hand to anyone starting out, in the order they cost me money.
If you want this voice today, prompt a frontier model. One line of instruction got me cleaner prose than three weeks of training did. If you want a fine-tune that genuinely wins on voice, I think it is findable and I did not find it: a corpus weighted hard toward long-form essayists, an explicit penalty against list structure, and a judge that scores the shape of the prose instead of which words show up in it. Whether that works, I don't know.
Claude told me I would lose, and Claude was right. Being told cost me nothing and taught me nothing. Losing cost me $113 and three weeks, and I now understand with my hands, rather than in the abstract, why training is hard and why the people who are good at it are worth listening to closely. The models will keep getting better at telling us things. We will keep needing to go and find out.
The model is public if you want to poke at it, and the full list of what went wrong is much longer than this post. If you have run this experiment and found the corpus that wins on voice, I would like to hear about it.
No comments :
Post a Comment