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 Formula Could Have Failed

By Joshua Bolick, Awareness Software Group ~12 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: What actually makes training cheaper? – One thing has worked: stopping part of the training early saved about 7% with no loss of quality. Everything else tested has been matched by a simpler or cheaper method -- and in two cases the clever method was only winning because it was quietly being given more.

In plain English

What we asked. When a new model starts from a piece of an older one, the older one has to be trained first, and there is a best amount: too little and it does not help, too much and it costs more than it saves. We had a simple formula for the best amount, about 1.4 times as long as a new model takes to learn the task on its own. But the earlier test of it could not have failed, because of how the experiment had been set up.

What we found. So we built a test where it could fail: a new task, and training lengths chosen without reference to the formula, where the formula and its only rival pointed at different answers. The formula said 200 steps; the rival said 120. The best length was 200, where it took fewer than four new models to repay the older one's training, against more than six at 120.

Why it matters. So the formula has earned some trust, and finding the right training length now takes one short trial run and a multiplication instead of a sweep. The lesson worth copying: a prediction that matches every case you have tried is worth nothing until you have tried a case where it could have been wrong.

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. The task was chosen by calibration on throwaway seeds, and the calibration, the exact-rung decision rule and the pilot were committed (7affbb4) before any measured run.

Program v2 Bucket R, item R12. Decisive computation: analysis/unscaled_ladder.py. Output: analysis/unscaled_ladder.json. Reproduce with python analysis/unscaled_ladder.py (about forty minutes on a throttled laptop CPU); --reuse re-derives every endpoint from the saved series.

The question

R5 found the donor break-even U-shaped in donor length, so the cheapest donor is not the best one and finding the best costs a sweep. R9 found one constant, 1.40 times the receiver control's transition time, that predicted the optimum on both committed ladders exactly -- and could not have failed: R5's ladder is P8's scaled by that same quantity, so "a fixed multiple of the transition" and "the same rung index" landed 0.037 of a rung apart. R14 counted it as one of the archive's three known designs that decided their own answer.

**How long to train the model you borrow from: a formula, tested fairly.** A new model can start from a piece of an older one, and the older one has to be trained first. Train it too briefly and it does not help; too long and it costs more than it saves. This measures the payback at each length, on a task chosen so two rival rules for the best length point at different answers. The best length is 200 steps, exactly where the formula put it: about 1.4 times as long as a new model takes to learn the task alone. Earlier tests of the formula could not have caught it being wrong; this one could, and it was right. Finding the best length now takes one short run and a multiplication instead of a sweep.

The fix is a ladder the predictor had no hand in. P8's absolute rungs, 20 to 400, unscaled, on a task whose transition sits between the two existing ladders'. Calibrated on throwaway seeds before the run: dispatch-copy-wide-vocab reaches the task criterion at about 150 steps. There the multiple predicts the 200 rung and the rung index predicts 120.

Decision rule, fixed before execution: the optimum is the rung with the lowest break-even (receivers needed for an own-donor to pay for itself, R5's definition, only where the head start excludes zero). The hypothesis whose rung equals the measured optimum wins; if neither, both lose. Because the two predictions are adjacent rungs, "within one rung" -- R9's tolerance -- would have accepted either, and is not used.

Kill test (the item's own): the predictions do not separate by at least one rung. Checked in code from the measured control, before the optimum is read.

The predictions separated

The measured control reaches the criterion at 153.3 [143.6, 163.1] steps. The multiple predicts 1.40 x 153.3 = 215, nearest rung 200; the rung index predicts 120. One rung apart, so the kill test does not fire and the ladder can decide.

Result: the optimum is at 200 -- the multiple wins

Donor length	Receiver head start	Receivers to pay back	Donor accuracy
`20`	`+22.5` `[+11.6, +33.4]`	no saving	`0.082`
`40`	`+25.8` `[+12.8, +38.9]`	no saving	`0.150`
`60`	`+17.5` `[+4.8, +30.2]`	no saving	`0.186`
`80`	`+6.7` `[+0.3, +13.0]`	no saving	`0.226`
`120` (rung index)	`-19.2` `[-30.4, -8.0]`	`6.3`	`0.298`
`200` (the multiple)	`-53.3` `[-65.1, -41.5]`	`3.8`	`0.388`
`400`	`-73.3` `[-83.7, -63.0]`	`5.5`	`0.493`

The best donor is 200 steps long, exactly where 1.40 x the control's transition put it, and the rung-index hypothesis loses by a factor of 1.7 in break-even. The U-shape reproduces on a third task: short donors delay the receiver, the optimum is in the middle, and the longest donor, though it gives the largest head start, costs more than it saves relative to 200. Short donors hurt here, as they did in P8: a donor must get past its own transition before it helps at all.

What this changes

R9's formula now carries evidence. On two scaled ladders it could not fail; on this unscaled one it could, and it picked the right rung against the one rival that had matched it everywhere else. The practical rule for this programme's donor economics:

Train the donor for about 1.4 times as long as a receiver takes to learn the task on its own. Measuring that takes one control run, which the receiver needs anyway.

The seven-rung sweep that R5 needed to find its optimum becomes one control run and a multiplication.

What stands

The kill test does not fire: the two predictions separate by a rung on an unscaled ladder.
The multiple wins: the optimum is at 200, the rung 1.40 x 153.3 predicts; the rung index's 120 pays back at 6.3 receivers against 3.8.
R9 is resolved by banner: its predictor has now passed a test it could have failed.
The U-shape holds on a third task.

Limits

The optimum is read from mean head starts. The break-even difference between the 200 and 400 rungs (3.8 against 5.5) has no interval of its own; the head starts at both exclude zero by a wide margin, and the ranking between them is a ratio of means.
One test of an exact-rung prediction on adjacent rungs. A single success is evidence, not a calibration: the constant 1.40 was fitted on two ladders and confirmed on one.
Six seeds, one width, one learning rate. The constant is stated for this setting.

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.
calibrated: A model is calibrated when its confidence matches how often it is right: answers it gives with 90% confidence should be right about 90% of the time.
calibration: Working out an instrument's settings from runs whose answer you already know, so it can be used on a run whose answer you do not.
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.