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 Quarter Is Free

By Joshua Bolick, Awareness Software Group ~13 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. The strangest result we have published is that copying a finished model into a fresh one can be much worse than doing nothing, if the two jobs are unrelated. It nearly doubles the time to learn, which is far worse than starting from scratch. Three of our results have now been built on top of that and none of them asked why it happens.

What we found. So we varied how much gets copied. A quarter of the wrong model costs nothing measurable at all. Half costs a little. The last quarter alone costs seventy steps. The damage is not proportional to the dose; there is a tolerance and then it accelerates.

Why it matters. The most informative test was to take the same weights and shuffle them into a random arrangement, keeping every number exactly but destroying the pattern. That removes about two thirds of the damage. So the problem really is the arrangement the wrong model learned, not that its numbers are unusual, and the remaining third splits into two smaller effects we can also name. Grafting later in training doubles the harm again, which fits the same story: there is more of the model's own work to destroy. We are careful in the record to say this is an explanation that fits two predictions rather than a mechanism we watched happen.

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, 7 arms over 6 seeds, no GPU, no cost.

Program v2 Bucket R, item R4. Decisive computation: analysis/why_the_wrong_donor_hurts.py. Output: analysis/why_the_wrong_donor_hurts.json. Reproduce with python analysis/why_the_wrong_donor_hurts.py in about fifteen minutes on a throttled laptop CPU; --reuse re-derives every endpoint.

The question

Q4 found that a donor converged on an unrelated task delays the receiver by +145.0 steps, 86% of its entire time to the transition and 6.7x worse than a random initialisation. Q14 reproduced it across a matrix and R2 showed it is large enough to make an unscreened donor worse than no donor at all.

**A quarter of the wrong model is free. The last quarter costs seventy steps.** Copying a finished model into a fresh one helps when the two jobs are related and hurts badly when they are not. Here we vary how much gets copied, from a quarter to all of it. The dashed lines are two comparisons: the same weights with their values shuffled into a random arrangement, and a model that was never trained at all. The damage is not proportional to the dose. A quarter costs nothing measurable, and then the curve accelerates. The shuffled line matters most: rearranging the same numbers removes about two thirds of the harm, which says the problem is the arrangement the wrong model learned rather than anything about the size of its numbers. That also means a small or blended copy is not automatically the bad idea a full one is.

Three records had built on it and none had asked why.

The hypothesis on trial is unlearning: the receiver must undo the donor's structure before building its own, so the cost is proportional to how much wrong structure it was given.

Kill test, fixed before execution: the harm is flat in graft fraction.

Prior recorded in advance: harm rises with fraction, and shuffling removes most but not all of it. Both halves were right, which is worth saying because the recorded prior has been wrong twice in the last week.

Anchor: Q4's harmful cell reproduces at +145.0, exactly, under Q4's own endpoint.

Result: the kill test does not fire, and the curve is not a line

Steps added against the receiver's own control, which reaches the criterion at 169.2.

Arm	Steps added
a quarter of the donor	`+0.0` `[-3.3, +3.3]`
half	`+19.2` `[+7.0, +31.3]`
three quarters	`+74.2` `[+61.1, +87.2]`
the whole donor	`+145.0` `[+118.9, +171.1]`

A quarter of the donor is free. The interval sits on zero. Going from a half to the whole donor multiplies the harm by 7.6x, and the last quarter alone costs +70.8 steps.

This is not the linear cost that "undo what you were given" would predict. There is a tolerance and then a cliff: a receiver absorbs some foreign structure without paying, and past that the cost accelerates.

What the harm is made of

The discriminating arm is the shuffled donor: the same weights with their entries permuted within each tensor. Every value is preserved, so norm, distribution and scale are identical by construction, and only the arrangement is destroyed.

	Steps added	Share of the `+145.0`
the whole donor	`+145.0`
the same values, shuffled	`+49.2` `[+39.5, +58.8]`
an untrained initialisation	`+21.7` `[+7.3, +36.0]`

Which decomposes it three ways:

+95.8, about two thirds, is the donor's arrangement. Shuffling removes it. This is structure that is actively wrong for the receiver's task, not weights that happen to be unusual.
+27.5, about a fifth, is the value distribution, the gap between shuffled weights and a fresh initialisation. Converged weights are differently distributed from initialised ones and that costs something on its own.
+21.7, about a seventh, is being handed somebody else's weights at all, which P1 measured and this reproduces.

Q4, Q14 and R2's framing survives. Had shuffling removed all of the harm, "converged on the wrong task" would have been too specific and all three would have needed weakening to "converged at all". It removes two thirds, so the specific framing is right and the record says what the other third is.

Grafting later hurts more, which is the second axis

Moving the graft from step 30 to step 120, at full strength, takes the harm from +145.0 to +287.5 [+256.2, +318.8]. The receiver has built more by then, so more is destroyed.

Both axes point the same way. The unlearning story survives its own test on dose and on timing, which is more than most explanations in this programme have managed.

What stands

The harm scales with dose and the scaling is superlinear. A quarter is free; the last quarter costs +70.8 steps.
Two thirds of it is the donor's arrangement, separated from its value distribution on the same weights.
Grafting later doubles it, from +145.0 to +287.5.
Unlearning is supported on both axes, and this is the first mechanism offered for the most surprising result the programme has produced.
There is a usable rule in the dose curve, if it holds: a small amount of a wrong donor costs nothing, so a blended or partial graft is not automatically a bad idea the way a full one is.

Limits

One donor, one receiver, one pair. running-count into dispatch-copy-8lag, chosen because it is the cell Q4 measured. Whether the shape holds for other harmful pairs is untested and is the obvious next question.
One width (48), one learning rate, six seeds, all inherited from Q4.
"Unlearning" is a story that fits, not a mechanism observed. Nothing here watches the receiver undo anything. Two predictions were derived and both held, which raises the story's standing and does not establish it. A version that measured the weights moving back toward their own trajectory would.
The late graft has a confound worth naming. Grafting at 120 on a run that transitions at 169.2 lands 71% of the way to the transition, so it destroys more progress and leaves less run to recover in. Those cannot be separated here, and the doubling should be read as consistent with unlearning rather than as a clean measurement of it.
The quarter arm blends against the receiver's own weights at the graft step, not against a fresh initialisation. That is the right comparison for a dose curve and it does mean "a quarter of the donor" is also "three quarters of what the receiver already had".

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.

confound: A second explanation you did not control for. If bigger models both learn faster and score higher, then 'fast learners score higher' may be entirely about size and not about speed.
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.