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

Too Weak to Learn Costs Nothing

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: What actually happens at the moment a model learns? – It builds machinery rather than selecting it, working through the task in a reproducible order and trying a simpler wrong rule on the way. The visible training curve cannot tell you which is happening.

In plain English

What we asked. We know that teaching a small model an extra skill it does not need slows its other skills. We wanted to know why: because the extra skill takes up room once learned, or because teaching it pulls on the model's settings every step. So we taught it only lightly, on one example in ten, expecting it to still be learned but pull far less.

What we found. It was never learned at all, on any run. A skill that is never learned takes up no room, so the test could not tell the two explanations apart. What we can say is that a skill taught this lightly, and not learned, slowed nothing.

Why it matters. We had written a rule that same morning: before a long run, do one short trial to check the thing the test depends on actually happens. We followed it once and skipped it here. The next attempt starts with that trial.

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, 7 training runs of 3000 steps. The design, precondition and kill test were committed (b0f271a, lint and backlog fixed in fd16623) before any run.

Program v2 Bucket O, item O22. Decisive computation: analysis/competition_mechanism.py. Output: analysis/competition_mechanism.json. Reproduce with python analysis/competition_mechanism.py (about forty minutes on a throttled laptop CPU); --reuse re-derives every endpoint from the saved series.

The question

O20 and O21 found that teaching a model one skill nothing needs (G) slows its later skills, and that doubling the width barely changes the share. Two readings remain: capacity (G occupies room once learned) and gradient (G's loss pulls on the shared weights every step). O22 tried to separate them by teaching G on only 10% of sequences -- still learned, the design assumed, so still occupying room, with a tenth of the pull.

**A skill taught too weakly to be learned costs the other skills nothing.** Three skills learned by the same small models under three conditions: an unneeded skill G taught on every sequence, taught on only one sequence in ten, and never taught. At one in ten, G itself was never learned. Taught on one sequence in ten, G is never learned and slows nothing: the other skills arrive as if G were absent. The test we designed needed G to be learned, so it could not separate the two explanations; a check we skipped would have caught that.

Kill test, fixed before execution: the share of F's full slowdown that remains under weak teaching has an interval including 1 (capacity).

Precondition, fixed before execution: G acquired on at least four of six seeds under weak teaching.

Anchor, in code: the first seed, with the batch wrapper installed at 100%, reproduces O17's committed series exactly. It does.

Result: the precondition fails -- `G` was never learned

Under 10% teaching, G is acquired on 0 of 6 seeds within 3000 steps. A skill that is never learned occupies no room, so the two readings make the same prediction and the kill test is not scored, as fixed in advance.

This was avoidable. The standing rule written earlier the same day -- size the budget against the slowest thing the test needs, with a throwaway run first -- applied here: the test needed G to be learned under weak teaching, and one throwaway run would have shown it is not. The rule was followed for O17 and skipped for O22.

What the runs show, described and not tested

Acquisition step, mean over seeds:

Skill	`G` fully taught (O17)	`G` taught on `10%`	`G` never taught (O20)
`C`	`1858`	`1073`	`1203`
`F`	`2313`	`1335`	`1433`
`E`	`2730` (4)	`1867` (4)	`1868`

A skill taught too weakly to be learned costs the other skills nothing: with G on 10% of sequences, C, F and E arrive as they do with no G at all. So the competition O20 found is not triggered by the mere presence of a G loss term at a tenth of its weight. That is consistent with both readings -- the pull is small, and nothing is learned to occupy room -- and so it separates nothing.

What stands

The precondition fails; the kill test is not scored. G is not learned at 10% teaching.
Descriptively, a weakly taught, unlearned skill costs the others nothing.
Generates O23: the same design at a teaching share at which G is learned, found on a throwaway seed before the pilot is committed.

Limits

One teaching share, chosen without the throwaway check the standing rule requires.
The remaining-share numbers printed by the script are not interpreted: with G unlearned they are ratios of noise to a large denominator.

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.

gradient: The direction and amount by which each of a model's internal numbers should change to do slightly better. Training is repeatedly following it.
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.
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 precondition fails -- G was never learned

What the runs show, described and not tested

What stands

Limits

Terms on this page

Read next

Get new results as we publish them

Want this measured on your data?

Result: the precondition fails -- `G` was never learned