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

One More Skill Slows the Rest

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. Last time we found that holding back a skill nothing else needed made a small model learn its other skills sooner, and guessed that skills compete for the model's limited room. That comparison had loose ends, so we ran a clean one: the same models, the same inputs, with one difference, whether that unneeded skill was ever taught.

What we found. When it was never taught, the later skills arrived hundreds of steps sooner. Teaching a small model one extra thing it does not need slows down everything it learns afterwards.

Why it matters. So how long a skill takes to learn is not a fact about that skill alone: it depends on everything else the model is being asked to learn at the same time.

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 plus O17's committed controls. The design and kill test were committed (fbbf16c); the kill-test expression was corrected to match its stated wording (c7baa85) before any run.

Program v2 Bucket O, item O20. Decisive computation: analysis/skill_competition.py. Output: analysis/skill_competition.json. Reproduce with python analysis/skill_competition.py (about thirty minutes on a throttled laptop CPU); --reuse re-derives every endpoint from the saved series, and was checked to print the identical summary.

The question

O17 withheld G -- a look-up nothing depends on -- for 600 steps, and the other skills arrived hundreds of steps sooner. Across experiments, adding G had also nearly doubled every composite's time against O12's control, but that comparison changed seeds and sequence length as well. The reading that fits is competition: every supervised skill draws on a shared, limited budget in a width-64 model. O20 makes the comparison paired.

**Teaching one skill that nothing needs slows every skill that comes after it.** The same six models trained twice on identical inputs: once taught every skill including G, a look-up no other skill uses (first bar of each pair), and once never taught G (second bar). The first skills are learned before the two runs differ. Without G, the later skills arrive hundreds of steps sooner. Every skill a small model is taught takes a share of a limited budget, even one that nothing else uses, so how long a skill takes depends on everything else in the task.

One new arm on O17's task and seeds: G-never, in which G's answers are unsupervised for the whole run. The inputs, sequence, query markers and every other skill's supervision are identical to O17's control; the only difference is whether a skill nothing needs is being taught.

Kill test, fixed before execution: G-never speeds neither C nor F (neither paired interval lies wholly below zero).

Anchor, in code: the first seed's control, re-run, reproduces O17's committed series exactly. It does; the other five controls are read from O17's output.

The kill-test correction, disclosed. The first committed expression also required each interval's lower bound to be at or below zero, which would have scored a slowed C or F as "does not fire". It was replaced by the stated test -- fire unless an interval lies wholly below zero -- in a separate commit before anything ran.

Result: the kill test does not fire

Acquisition step, mean over seeds that acquired it, and the paired change:

Skill	Control (teaching `G`)	`G-never`	Paired change
`B`	`118`	`113`	`-5.0` `[-16.0, +6.0]`
`D`	`667`	`400`	`-266.7` `[-371.6, -161.8]`
`A`	`830`	`490`	`-340.0` `[-386.9, -293.1]`
`C`	`1858`	`1203`	`-655.0` `[-1449.1, +139.1]`
`F`	`2313`	`1433`	`-880.0` `[-1537.2, -222.8]`
`E`	`2730` (4)	`1868`	`-755.0` `[-1293.4, -216.6]` (4 seeds)

Not teaching G at all speeds F by 880 steps and E by 755, intervals clear of zero; C moves by -655 with an interval that reaches +139. The kill test does not fire.

A and D match O17 exactly, and should. Both are acquired before step 600 in both arms, and up to step 600 a run that withholds G for 600 steps and a run that withholds it for all 3000 are the same run. They carry no new information; the skills acquired after step 600 do.

Reading it

With the inputs identical, teaching one extra skill that nothing needs makes the other composites take hundreds of steps longer. That is competition for a shared budget -- capacity, gradient signal, or both; this record does not separate them. It also fits the cross-experiment comparison: G-never acquires C at 1203 and F at 1433, close to O12's control without G (1038, 1290) although seeds and sequence length differ.

It re-reads O12 too. There, withholding a prerequisite delayed most skills. If skills compete, freeing a prerequisite's share should help; it did not, so a prerequisite's absence costs more than its share of the budget returns. Why is open.

What stands

Teaching a skill nothing needs slows the others: without it, F -880.0 [-1537.2, -222.8] and E -755.0 [-1293.4, -216.6], on paired seeds with identical inputs.
Competition for a shared budget is supported, not yet decomposed.
Implication for this programme's multi-skill tasks: every skill in a task affects the timing of every other. Acquisition times are properties of the whole task, not of a skill.

Limits

One width (64); competition for capacity should ease as width grows, which is untested.
C's interval includes zero on six seeds with spread of hundreds of steps.
Capacity or gradient: withholding supervision removes both G's loss term and its demand on the model; which matters is not separated.

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.