How do you know a fine-tuned AI model is actually better?

You measure it. Hold out a portion of the client's real examples so the model never sees them during training, score the plain base model on those held-out questions first, fine-tune, then score again on the identical questions with identical scoring. The difference is the lift. Because the only thing that changed is the fine-tune, the improvement is attributable to it.

Does fine-tuning actually improve accuracy, or is the base model already good enough?

It depends on the task and, critically, on how much data you have. On a public benchmark of 510 lawyer-annotated commercial contracts, a base Qwen3-4B extracted short contract fields at 62.2% accuracy. Fine-tuned on 396 examples it reached 83.5-85.4%, a lift of 21.3 to 23.2 points across three runs. The same model trained on only 48 examples produced a lift inside the noise floor, so the amount of data was the deciding factor rather than the model's ceiling.

How much of our own data do we need to fine-tune a model?

Fewer examples than most people expect, but more than a handful. In our contract-extraction measurements, 48 training examples produced no measurable improvement while roughly 400 produced a 22-point gain. The honest answer for any specific case is that it depends on how varied and how consistent the examples are, which is what a paid readiness assessment measures on the client's own corpus.

Why can't you just use a public AI benchmark score?

Public benchmarks measure general ability on standardised tests, not your work. A model that scores well in the abstract can still be mediocre at reading your contracts or following your policies. Public benchmarks are also increasingly targeted during training, so a high score can reflect test preparation as much as capability. The only benchmark that answers your question is built from your own tasks.

Research record

One Hundred Steps Is Enough

By Joshua Bolick, Awareness Software Group ~9 min read September 2026

Ongoing research. This is an experimental result from active work, not a settled conclusion. The numbers are what we measured and the method is described so you can judge it, but the programme is still running and later experiments may revise what it means.

A note on the language in these records. This is a working laboratory notebook for research into training AI models more cheaply and efficiently, so you will read that an approach did not work, that a result did not hold up, or that one method was worse than another. That is the research doing its job, not a verdict on the engineering we deliver to clients. Ruling an approach out is how the search narrows, and these are the pages that teach us the most: nearly every technique we now rely on came from understanding why something else fell short. Testing our own ideas at least as hard as anyone else's is the point of publishing them. More about this programme and why we run it.

Part of a bigger question: Is the task we are studying actually hard? – Often it is not. A rule from 1990 with no parameters beats the trained model on the task most of these results were measured on, and what an intervention costs is set by the task's own structure.

In plain English

What we asked. A short warm-up on easier versions of a task makes a small model much better at the hard version. We had shown a warm-up of 6 percent of training was enough. How short can it go?

What we found. A hundred steps, under 2 percent of training, gave the whole benefit: 82 percent on the hard task against 58, on every run. Fifty steps did much less. And the benefit did not appear during the warm-up; it showed up thousands of steps later.

Why it matters. The tiny burst seems to change which way training goes rather than teach much directly. If you are going to try this, it is almost free to try.

The rest of this page is the technical record: the design, every number, and the limits. It is written for a reviewer, and you do not need it to have understood the result above.

New here? How to read a research record

Start at the verdict. Every record states, before the experiment was run, what result would have made us abandon the idea. That is the "kill test". Then it says whether the test fired. Nothing gets reinterpreted after the fact.
Numbers in square brackets are uncertainty. 23.4 [18.1, 28.7] means our best estimate is 23.4 and the true value is probably somewhere in that range. If a range includes zero, we cannot claim an effect.
Results that rule an idea out are kept. Roughly half of what is published here says an approach did not work, including plenty of our own. Those pages are the output, not a shortfall: knowing which direction is a dead end is what lets the next experiment go somewhere better, and most of what we now rely on came out of understanding why something else fell short. Work that only publishes what worked is not measuring anything.
Read the Limits section. Every record ends with what it does not show. It is the most honest part of any experiment and usually the shortest.
Pro tip: the figures near the top are designed to carry the result on their own. If you read nothing else, read the caption under each one, which says what it shows and what to take from it.

EXPLORATORY. Not a preregistered study. Local CPU, 48 training runs of 6000 steps. The design and the kill test were committed (46c9605) before any run.

Program v2 Bucket A, item A23. Decisive computation: analysis/warmup_floor.py. Output: analysis/warmup_floor.json.

The question

A20 found a random mix of easier copy lags for only the first 375 of 6000 steps lifts final lag-8 accuracy by +0.258, as much as any longer warm-up. The floor was not found. How short can it be?

**One hundred steps of easier examples is enough.** Every model trains for the same 6,000 steps. The grey bar trains on the hard task throughout; the others start with a very short burst of easier versions of it, from 25 to 375 steps, then switch to the hard task. Error bars are 95% confidence intervals. *Somewhere between 50 and 100 steps the benefit switches on. A hundred steps, under 2% of training, lifts the final result from about 58% to 82% on every run, as much as any longer warm-up.*

Design: A20's code, task, rate 0.01 and seeds (J8's donor seeds). Warm-ups of 25, 50, 100 and 200 steps (lags 2, 4, 6 drawn at random each step), then lag 8 to step 6000; the 0- and 375-step arms re-run as anchors. Kill test, fixed before execution: the 100-step warm-up's gain over fixed includes zero. Anchor, in code: the 0- and 375-step arms reproduce A20 exactly -- held on all eight seeds.

Results

Warm-up (steps of `6000`)	Final lag-`8` accuracy	Minus fixed, paired	Seeds above fixed
`0` (fixed)	`0.575`	--	--
`25`	`0.658`	`+0.084` `[-0.011, +0.178]`	`6` of `8`
`50`	`0.668`	`+0.093` `[+0.008, +0.178]`	`7` of `8`
`100`	`0.817`	`+0.242` `[+0.163, +0.321]`	`8` of `8`
`200`	`0.798`	`+0.223` `[+0.121, +0.325]`	`8` of `8`
`375` (A20)	`0.833`	`+0.258` `[+0.159, +0.357]`	`8` of `8`

The kill test does not fire. A warm-up of 100 steps -- under 2% of the run -- gives the whole gain. Between 50 and 100 steps the effect switches on: 50 steps is only just distinguishable from nothing (+0.093), and it is uneven -- one seed ends at 0.24, the rest between 0.66 and 0.78 -- while at 100 every seed ends between 0.77 and 0.93.

The gain appears late, not during the warm-up. At step 1000 the arms differ by a few points (0.369 fixed, 0.419 with 100 warm-up steps); at step 2000 by 0.09; by step 6000 by 0.24. A hundred steps at the start changes where training ends up, not how far it has got by the time the burst is over.

What it says

With A20 and A21: about a hundred steps of easier versions of the target, at the very start, then the target gives the full effect on this task. That is too little training to have learned the easy lags to any depth, and the gain shows up thousands of steps later, so the reading this favours is that the burst decides which solution the model goes on to find -- an effect on the path, not a stock of knowledge. The obvious test of that reading is whether the same 100 steps work later in training: A26.

What stands

A23: kill test does not fire. 100 steps give +0.242 [+0.163, +0.321], on every seed.
The threshold lies between 50 and 100 steps on this task at rate 0.01.
The gap opens long after the burst ends.

Limits

One task, one width (48), one rate; the threshold in steps is likely to depend on all three.
The 25- and 50-step arms are noisy; their intervals are wide and one seed dominates the 50-step spread.

Terms on this page

Every piece of vocabulary this record uses, in plain language. Generated from the text above, so it cannot drift out of step with it.

accuracy: The fraction of answers a model gets right on questions it was not trained on.
kill test: A condition written down before running the experiment that says what result would make us abandon the idea. Fixing it in advance is what stops a disappointing result being reinterpreted as an encouraging one.
seed: The number that fixes all the randomness in a training run. Same seed, same run. Running several seeds is how you tell a real effect from a lucky one.
width: How many internal numbers a model uses at each layer. The usual way we vary model size in these experiments.