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

With Both Tuned, the Warm-Up Is Within Noise

By Joshua Bolick, Awareness Software Group ~10 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. After finding that our warm-up results had been measured against a poorly tuned baseline, we tried five learning rates for both plain training and training with the warm-up.

What we found. Each at its own best setting, the warm-up came out 3 points ahead, well within noise. What it really did was move the best setting: plain training liked small steps, the warm-up version liked large ones.

Why it matters. Tune both the new method and the old one before comparing them. Here that erased the result.

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 new training runs of 6000 steps plus 32 re-used. The design and the kill test were committed (989e006) before any run.

Program v2 Bucket A, item A31. Decisive computation: analysis/warmup_tuned.py. Output: analysis/warmup_tuned.json.

The question

A29 found every copy-task warm-up record compared against an untuned rate: at 0.005 the plain run reaches 0.763 and the 1500-step warm-up adds nothing. With each arm at its own best rate, does the warm-up beat the plain run?

**Warm-ups do not raise the best result; they move the best learning rate.** Each point is eight training runs of 6,000 steps. Plain training on the hard task does best at a small learning rate and falls apart at large ones. Both warm-ups do worse than plain training at small rates and better at large ones. Compared at each one's own best setting, the long easy-task warm-up is no better than plain training. The short burst beats it only at the largest rate, which is what a standard learning-rate warm-up also achieves.

Design: A15's task and seeds (J8's donor seeds), width 48, 6000 steps. Fixed lag 8 and A15's 1500-step shuffled warm-up at rates 0.002, 0.003, 0.007 (new), 0.005 (A29) and 0.01 (A15). Each arm's best rate by mean final accuracy; both tuned on the same seeds. Kill test, fixed before execution: best shuffled minus best fixed, paired, includes zero or lies below it; a best rate at a grid edge is reported as not found. Anchor, in code: fixed at 0.005 on the first seed reproduces A29 exactly -- held.

Results

Rate	Fixed lag `8`	Shuffled warm-up, then lag `8`
`0.002`	`0.694` `[0.673, 0.715]`	`0.594` `[0.544, 0.643]`
`0.003`	`0.768` `[0.728, 0.808]`	`0.628` `[0.589, 0.667]`
`0.005`	`0.763` `[0.715, 0.812]`	`0.729` `[0.696, 0.762]`
`0.007`	`0.735` `[0.686, 0.785]`	`0.796` `[0.764, 0.828]`
`0.01`	`0.575` `[0.492, 0.657]`	`0.790` `[0.765, 0.815]`

The kill test fires. Each arm's best rate is inside the grid (fixed 0.003, shuffled 0.007), so both are found. Best shuffled minus best fixed is +0.028 [-0.025, +0.080]: with both tuned, the warm-up's advantage is within noise. At 0.01 it was +0.215; almost all of that was the plain run's poor rate.

The two arms prefer different rates. The plain run is best at 0.003 and falls off sharply above 0.005; the warm-up arm is best at 0.007-0.01 and is worse than the plain run at every rate up to 0.005. The warm-up lets training use a larger rate without the damage a large rate otherwise does early -- which is what a learning-rate warm-up is for, and A33 found a 375-step rate warm-up does it better.

What it says

Against the best constant rate, the 1500-step data warm-up is not an efficiency gain. The thread's effect is real and specific at a high rate, and it is the effect of protecting the first steps from that rate. Two questions remain, and they are the honest end of the thread:

A34 -- A30's 100-step lag-5 burst, the strongest data warm-up, at the plain run's best rate and its neighbours.
A35 -- whether a data burst adds anything on top of a learning-rate warm-up, which is the baseline every modern recipe already uses; and the rate warm-up's own peak tuned, since at 0.01 it reached 0.894, above the best constant rate.

What stands

A31: kill test fires. Best shuffled (0.796 at 0.007) minus best fixed (0.768 at 0.003): +0.028 [-0.025, +0.080].
The warm-up shifts the best rate upward (0.003 to 0.007) rather than raising what training reaches.

Limits

Selection on the same eight seeds that measure the difference; the bias is shared by both arms but not removed.
Five rates, constant schedules only.

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.
baseline: The thing you compare against. A result without one is not a result.
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.
learning rate: How big a step training takes each time it updates the model. Too small and nothing happens; too big and it never settles.
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.