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 of One Easier Version

By Joshua Bolick, Awareness Software Group ~9 min read August 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. Our experiments on warming a model up with easier examples had narrowed the recipe down: brief, at the very start, and a version of the task easier by a clear margin. We tried the simplest form of it: 100 steps of one easier version.

What we found. It was the strongest warm-up yet: 89 percent on the hard task against 58 without it, on every run, and above the 76 percent that plain training reached at a better learning rate.

Why it matters. That last comparison is across two settings neither of which we had tuned, so it is a hint, not a result. The next experiment tunes plain training properly and runs this warm-up against it.

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, 24 training runs of 6000 steps. The design and the kill test were committed (53ba049) before any run, and before A29 found this thread's rate (0.01) untuned. Read with it.

Program v2 Bucket A, item A30. Decisive computation: analysis/warmup_minimal.py. Output: analysis/warmup_minimal.json. Descriptive post-hoc: analysis/warmup_minimal_posthoc.py and its output.

The question

A23: 100 steps of the lag 2/4/6 mix at the start give the whole gain. A26 and A28: only in the first hundred steps. A27: lag 5 alone is the best single-lag warm-up. Does the shortest, simplest version -- 100 steps of lag 5 -- work?

**A hundred steps of one easier version is the strongest warm-up yet.** Plain training on the hard task at two learning rates, and two brief warm-ups at the start: 100 steps of a random mix of easier versions, or 100 steps of just one, remembering 5 steps back. Error bars are 95% confidence intervals. The two rates were not tuned, so the comparison across them is indicative. *The single easier version for 100 steps reaches 89%, above both the mix (82%) and the better plain run (76%). Whether it still wins against properly tuned plain training is the next test.*

Design: A15's task, rate 0.01 and seeds (J8's donor seeds), 6000 steps. Steps 1-100 on lag 5 only, then lag 8; against fixed and A23's 100-step mix burst. Kill test, fixed before execution: the lag-5 burst minus fixed includes zero. Anchor, in code: fixed and the mix burst reproduce A23 exactly -- held on all eight seeds.

Results

Arm (rate `0.01`)	Final lag-`8` accuracy	Minus fixed, paired
fixed	`0.575` `[0.492, 0.657]`	--
`100` steps of the `2`/`4`/`6` mix (A23)	`0.817` `[0.772, 0.861]`	`+0.242` `[+0.163, +0.321]`
`100` steps of lag `5` only	`0.892` `[0.865, 0.920]`	`+0.318` `[+0.235, +0.401]`

The kill test does not fire. The lag-5 burst beats fixed on every seed, and beats the mix burst by +0.076 [+0.038, +0.114]. One hundred steps of one easier version, at the right distance from the target, is the strongest warm-up tried -- stronger than any 1500-step warm-up in the thread.

Against the better plain run (descriptive, post-hoc, not the kill test). A29 found the plain run reaches 0.763 at rate 0.005, which matched every earlier warm-up. The lag-5 burst at 0.01 beats that run by +0.129 [+0.082, +0.176], on all eight seeds. This compares across two rates, neither tuned for its arm, so it is a size, not a verdict. It is, though, the first warm-up in the thread to clear the better plain run at all.

What it says

The recipe distilled by A21-A28 -- brief, first, related, and easier by a margin -- is also the strongest version of it. Whether it survives a properly tuned plain run is the question the thread now turns on: A31 (running) tunes the plain run over five rates, and A34 runs the lag-5 burst at the plain run's best rate.

What stands

A30: kill test does not fire. 100 steps of lag 5 give +0.318 [+0.235, +0.401] at rate 0.01, every seed.
It beats the 100-step mix by +0.076 [+0.038, +0.114].
It clears the plain run at rate 0.005 by +0.129 (descriptive, cross-rate).

Limits

Rate 0.01, untuned (A29). A34 is the test that matters.
One task, one width, one burst length.

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.
post hoc: Worked out after the fact, rather than decided in advance. We report such checks separately and never let them decide a result, because it is far too easy to find a pattern once you already know the answer.
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.