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

A Higher Ceiling, Not a Head Start

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. We had found that a random mix of easier tasks early in training makes a small model better at a hard task. Two questions remained: is that only a head start the ordinary model would catch up on, and is it just that variety helps?

What we found. Given two and a half times longer, the ordinary models stalled around 64 percent while the mixed-start models climbed to 94: a genuinely better end point. And keeping the easy tasks mixed in for the whole run did worse than either. The benefit comes from a varied warm-up that then gives way to the real task.

Why it matters. In practice: a phase of varied, easier material before focusing on the target task can leave a model better off for good. Let the phase end.

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 15000 steps. The design and both kill tests were committed (97472f5) before any run.

Program v2 Bucket A, items A16 and A17. Decisive computation: analysis/lag_mixture_longer.py. Output: analysis/lag_mixture_longer.json.

The questions

A15 found an early random mix of easier copy lags raises final lag-8 accuracy by +0.215 at step 6000, and A18 found the same on a second task. Two readings remained:

**Mixing easy tasks first raises the ceiling; mixing them throughout does not.** Three ways to spend the same compute: the hard task throughout; a random mix of easier versions for the first 1,500 steps, then the hard task; or easy and hard versions mixed together for the whole run. Run 2.5 times longer than before. Given more time, the hard-only runs stall around 64% while the easy-mix-first runs climb to 94%: a better final solution, not just a faster start. Keeping the easy versions in the mix for the whole run does worse than either.

A16 -- a higher ceiling, or only a head start? Run A15's fixed and shuffled arms to step 15000. Kill test: shuffled minus fixed at 15000 includes zero (the fixed arm catches up).
A17 -- is it simply variety? A third arm mixes lags 2, 4, 6 and 8 at random for the whole run. Kill test: mixed-throughout minus fixed at step 6000 includes zero or lies below it.

A15's seeds; the first 6000 steps of the fixed and shuffled arms reproduce A15 exactly (checked in code on every seed).

Results

Lag-`8` accuracy	Step `6000`	Step `15000`
fixed lag `8`	`0.575` `[0.492, 0.657]`	`0.636` `[0.564, 0.709]`
shuffled early, then lag `8`	`0.790` `[0.765, 0.815]`	`0.940` `[0.918, 0.961]`
mixed lags throughout	`0.490` `[0.458, 0.523]`	`0.504` `[0.368, 0.639]`

A16: the kill test does not fire. At step 15000 the shuffled arm is +0.304 [+0.214, +0.393] ahead, a wider gap than at 6000. The fixed arm gains 6 points in 9000 more steps; the shuffled arm gains 15 and approaches 0.94. The early mixture raises the ceiling this model reaches, not just how fast it gets there.

A17: the kill test fires. Mixing all lags throughout ends -0.084 [-0.184, +0.016] below fixed at 6000 and -0.133 below at 15000. Variety for its own sake does not help; with the target lag only a quarter of the data, the model never specialises.

What the thread now says

With A14, A15 and A18: varied easier material early, then focused training on the hard task, lets a small model reach a better solution than it finds by training on the hard task alone -- on two task families, at matched compute. The order of the easy material does not matter (A14); the variety does, but only as a phase that ends (A17); and the gain is a higher ceiling (A16), not a head start.

What stands

A16: kill test does not fire. Shuffled minus fixed at step 15000: +0.304 [+0.214, +0.393].
A17: kill test fires. Mixing all lags throughout does not help (-0.084 at 6000, -0.133 at 15000).
The recipe is a phase: varied easy material first, then the target.

Limits

One width, tasks chosen near the edge of learnability; 15000 steps is still finite.

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.