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

The Second Head Start Survives Tuning

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: Which parts of a model actually matter? – The parts that move most are not the parts that matter, and no single component is required -- the model routes around every freeze. What a part is worth shows up only when you remove it.

In plain English

What we asked. Our strongest result is that a small model learns much sooner when handed part of another model that already learned the task. On our second task, the learning rate we had used turned out to be well below the best one, so we re-ran the result at the best rate.

What we found. It held. The head start got somewhat smaller, because the better rate itself made up some of the difference, but it still cut the time to learn by about half, and handing over an untrained part still slowed learning down.

Why it matters. The result has now survived a properly tuned baseline on both tasks we measured it on, which is the check most likely to have broken 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, six seeds, 60 training runs. The design, the graft rule and the kill test were committed (e80da4f) before any run; the tuned rate was read from O19's committed output.

Program v2 Bucket R, item R20. Decisive computation: analysis/dispatch_transfer_at_tuned_rate.py. Output: analysis/dispatch_transfer_at_tuned_rate.json. Reproduce with python analysis/dispatch_transfer_at_tuned_rate.py (about twenty-five minutes on a throttled laptop CPU); --reuse re-derives every endpoint from the saved runs, and was checked to print the identical summary.

The question

Q4 found that on dispatch-copy-8lag a same-task donor's converged recurrent weights remove 90.0 steps from a receiver's time to L7's criterion, at the house rate 0.005. O19's sweep then showed dispatch copy's tuned rate is 0.010 and the house rate 35.0 steps slower -- the largest rate exposure found for any claim outside J8's thread. R19 re-tested P1 on delayed copy, where the rate had 8.5 steps to give, and it survived. R20 repeats that on the task where the rate had four times as much to give.

**On the second task too, a trained donor's head start survives a tuned learning rate.** The second task, where our usual learning rate turned out to be well below the best. A model is handed part of another, already-trained model, or the same part from an untrained one, at the old and the tuned learning rate. The last bar compares a donor at the old rate with no donor at the tuned rate. The head start shrinks somewhat at the tuned rate, because the rate itself makes up part of it, but it still halves the time to learn, and the untrained part still slows learning. A trained donor is worth more than tuning the rate here too.

Every run is Q4's own loop (P1's run() on Q4's task), timed by L7's criterion as Q4 was: the same-task donor (800 steps at the same rate), control, recurrent and recurrent-untrained, at 0.005 and at 0.010. The graft moves proportionally by R19's rule (step 24 at 0.010); P1's step 30 is also run.

Kill test, fixed before execution: at 0.010 with the proportional graft, the recurrent head start's interval includes zero.

Anchor, in code: at 0.005, every seed's control and recurrent accuracy series equal Q4's committed ones exactly. All do.

Result: the kill test does not fire

Change in time to L7's criterion against each seed's own control (steps; negative is sooner):

	`0.005` (Q4, anchor)	`0.010` (tuned)
control	`169.2`	`134.2`
`recurrent` (same-task donor)	`-90.00` `[-99.39, -80.61]`	`-65.83` `[-74.87, -56.79]`
`recurrent`, graft at P1's step `30`		`-60.83` `[-72.53, -49.14]`
`recurrent-untrained`	`+21.67` `[+7.33, +36.01]`	`+22.50` `[+9.75, +35.25]`
head start as a share of the control's time	`53.2%`	`49.1%`

The head start survives, smaller in steps and nearly unchanged as a share. It falls by 24.2 steps -- close to what the rate itself saved the control (35.0), as if part of what the donor bought at the slow rate was speed the rate would have given anyway -- but it still halves the receiver's time. The untrained-donor control still delays by the same amount, so the benefit is still learned structure.

The practical comparison: a receiver given a donor at the old rate reaches the criterion 55.0 steps [-66.97, -43.03] sooner than a receiver alone at the tuned rate. Tuning buys 35 steps; a donor at the untuned rate still beats tuning alone by 55.

What stands

Q4 survives dispatch copy's tuned rate: -65.83 [-74.87, -56.79], 49% of the run against 53%.
Part of the old head start was rate: it shrank by 24 steps where the rate saved the control 35.
With R19, the transfer thread's head start now survives a tuned rate on both tasks it was measured on. R12's donor-length optimum, on the same substrate and rate, was not re-run.

Limits

One task, one width, six seeds; L7's criterion only.
The donor was retrained at each rate, so the comparison changes donor and receiver together, as a practitioner would.
R12's optimum is not re-tested; the head start at each donor length may shift differently.

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.
recurrent: A design that reads a sequence one item at a time, carrying memory forward. The main alternative is attention, which looks at everything at once.
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.

In plain English

The question

Result: the kill test does not fire

What stands

Limits

Terms on this page

Read next

Get new results as we publish them

Want this measured on your data?