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

Easier Data Adds Nothing to a Rate Warm-Up

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. Our easy-data warm-up turned out to do the same job as a standard learning-rate warm-up, which starts training with small steps. One question was left: does easy data add anything when training already has that standard warm-up?

What we found. No. With the standard warm-up at its best setting, adding our best easy-data warm-up changed the result by about one point, within noise. The best result in the whole study, 90 percent, came from the standard warm-up alone.

Why it matters. For this task, the ordinary recipe was already enough. The easy data only helped when the steps were too large for the ordinary warm-up to handle.

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, 64 training runs of 6000 steps. The design and the kill test were committed (1b962fb) before any run.

Program v2 Bucket A, item A35. Decisive computation: analysis/warmup_on_rate_warmup.py. Output: analysis/warmup_on_rate_warmup.json.

The question

A31, A33 and A34 found the copy-task data warm-ups only let training use a larger learning rate, and a plain 375-step learning-rate warm-up does that as well. The fair baseline for easier data is therefore a tuned learning-rate warm-up, which every modern recipe already has. Does the strongest data warm-up -- 100 steps of lag 5 -- add anything on top of one?

**On top of a learning-rate warm-up, easier data adds nothing.** Every run uses a standard learning-rate warm-up: small steps at first, rising to the peak over 375 steps. The dashed line also spends its first 100 steps on an easier version of the task, the strongest data warm-up we found. *At the best settings the two lines coincide at about 90%. The easier data only helps at a peak rate too large for the warm-up schedule alone. A standard warm-up schedule is all this task needs.*

Design: A15's task and seeds (J8's donor seeds), 6000 steps. A 375-step linear rate warm-up to peak 0.007, 0.01, 0.014 or 0.02, with or without the lag-5 burst in steps 1-100. Kill test, fixed before execution: at the rate warm-up's best peak, the combined arm minus the rate warm-up alone includes zero or lies below it; a best peak at a grid edge is reported as not found. Anchor, in code: the rate warm-up alone at 0.01 reproduces A33's runs exactly -- held on all eight seeds.

Results

Peak rate	Rate warm-up alone	Rate warm-up + lag-`5` burst
`0.007`	`0.903` `[0.860, 0.946]`	`0.892` `[0.876, 0.909]`
`0.01`	`0.894` `[0.871, 0.916]`	`0.894` `[0.852, 0.937]`
`0.014`	`0.739` `[0.561, 0.918]`	`0.880` `[0.840, 0.919]`
`0.02`	`0.337` `[0.140, 0.534]`	`0.324` `[0.197, 0.452]`

The kill test fires. At the rate warm-up's best peak the burst adds -0.011 [-0.045, +0.024]. The best peak (0.007) is at the grid's lower edge, so strictly it is not found; but 0.01 is within a hundredth, and the burst adds exactly nothing there either. The top of the curve is a plateau, and the burst does not raise it.

The same pattern as before, once more. At 0.014 -- past the plateau, where the rate warm-up alone starts to fail -- the burst helps (0.880 against 0.739): it again makes a larger rate usable. It does not make the best run better.

What the data-warm-up thread found, in the end

On the copy task, an early phase of easier data (A14-A30) does one thing: it protects the first steps of training from a learning rate that is otherwise too large. That is why it only works at the start (A26, A28), only at high rates (A29, A31, A34), and not on top of a learning-rate warm-up that already does the same job (A33, A35). The details found on the way -- the material must be related (A21), a sweet spot in how much easier (A27) -- describe how well easier data does that job. No data warm-up tested beats a tuned learning-rate warm-up. The modular-sum task is still open (A32, A36 running), where the mixture has so far held up at every rate.

What stands

A35: kill test fires. On top of a 375-step learning-rate warm-up at its best peak, the lag-5 burst adds -0.011 [-0.045, +0.024].
The best copy-task result in the thread (0.903) needs no special data: a learning-rate warm-up to 0.007.

Limits

One ramp length (375 steps), linear; one task. The best peak is at the grid edge (a plateau with 0.01).

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.

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.
plateau: A stretch of training where the model is not visibly improving. Often, though not always, followed by a sudden jump.
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.