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

Skills Compete

By Joshua Bolick, Awareness Software Group ~11 min read August 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, holding back one skill while a small model trained delayed the skills it normally learns afterwards, and we guessed the model learns skills one at a time, like a queue. To test the guess, we held back a skill that no other skill uses.

What we found. A queue predicts everything behind it should wait. Instead, everything else was learned hundreds of steps sooner, and only the held-back skill was delayed. Our guess was wrong. Adding that one unneeded skill had also made every other skill much slower to learn.

Why it matters. That looks like competition: a small model has limited room, and every skill it is asked to learn takes a share, even one nothing else needs. We will test that directly next.

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 fresh seeds, 12 training runs of 3000 steps. The design, precondition and kill test were committed (90cd477) before any run, with the budget checked on a throwaway seed first.

Program v2 Bucket O, item O17. Decisive computation: analysis/acquisition_queue.py. Output: analysis/acquisition_queue.json. Reproduce with python analysis/acquisition_queue.py (about fifty 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

O12 withheld skill A's supervision for 600 steps and found it delayed nearly every skill the model normally learns afterwards, including F, which does not need A. It read that, post hoc, as a queue: skills learned one after another, so holding one up holds up everything behind it. O17 tests the reading with a skill nothing depends on.

**Holding back a skill nothing needs makes every other skill arrive sooner.** Seven skills, one of which, G, is a simple look-up that no other skill uses. For the first 600 steps the model gets no feedback on G. Each bar is how much sooner or later each skill is learned than normal, averaged over six runs. If skills were learned in a queue, holding G back would delay everything after it. Instead everything else arrived hundreds of steps sooner, and only G was delayed. The skills look like they compete for the same limited room in the model.

O12's six skills plus a seventh, G, a look-up of position 9 that no composite reads. Two arms on six fresh seeds: control, and withhold-G (G's answers unsupervised for steps 1-600). Every run is O12's own train(), pointed at the seven-skill task.

Kill test, fixed before execution: withholding G delays neither C nor F -- both paired intervals include zero. The queue reading predicts that both are delayed.

Precondition: C, F and G acquired on at least four of six seeds in both arms. It passes.

Budget: 3000 steps, checked on throwaway seed 9999 before the pilot was committed (G at 410, F at 1900).

Result: the kill test does not fire -- in the direction neither reading predicted

Acquisition step, mean over the seeds that acquired it:

Arm	`B`	`G`	`D`	`A`	`C`	`F`	`E`
control	`118` (6)	`608` (6)	`667` (6)	`830` (6)	`1858` (6)	`2313` (6)	`2730` (4)
withhold-G	`113` (6)	`868` (6)	`400` (6)	`490` (6)	`1307` (6)	`1613` (6)	`2228` (6)

Change under withhold-G, paired against each seed's control:

Skill	Change (steps)
`G` (withheld)	`+260.0` `[+46.0, +474.0]`
`B`	`-5.0` `[-16.0, +6.0]`
`D`	`-266.7` `[-371.6, -161.8]`
`A`	`-340.0` `[-386.9, -293.1]`
`C`	`-551.7` `[-704.7, -398.6]`
`F`	`-700.0` `[-1189.2, -210.8]`
`E`	`-475.0` `[-1306.1, +356.1]` (4 seeds)

Neither interval includes zero, so the kill test as written does not fire -- but both lie below it. Withholding G does not delay the skills learned after it; it makes them arrive hundreds of steps sooner. The test was written to tell "delayed" (a queue) from "not delayed" (something A shares with its descendants). It did not anticipate "sped up", and that is the answer.

The queue reading is refuted in its simple form. A queue cannot make the items behind a held-up item move faster.

What the whole set of results now suggests

Two more observations, both post hoc:

Adding G slowed everything. This control, with G, acquires C at 1858 and F at 2313; O12's control, without G, acquired them at 1038 and 1290. The seeds and sequence length differ, so this is not a paired comparison, but the size is striking: one extra look-up that nothing uses nearly doubled the time to every composite.
Taking G away for a while gives much of that back (C -552, F -700).

Both point at competition: the width-64 model has a limited budget of capacity or gradient signal, and every supervised skill draws on it, whether or not anything depends on it. On that reading O12's delays under withhold-A are the harder case: withholding a prerequisite hurts broadly even though freeing its share of the budget should help. That contrast is worth a clean test. Generates O20: the same seeds with and without G, paired.

What stands

Withholding a skill nothing depends on speeds the others: C -551.7, F -700.0, A -340.0, D -266.7, intervals excluding zero. G itself is delayed +260.
The serial-queue reading of O12 is refuted.
Competition for a shared budget is the leading reading, and it is post hoc; O20 tests it.

Limits

The kill test's wording tested "delayed or not" and could not score a speed-up; the direction is reported here in full rather than left to the test's verdict.
The cross-experiment comparison (with and without G) uses different seeds and sequence lengths.
One width (64), one withholding length (600), six seeds; E acquired on four control seeds.

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.
post hoc: Worked out after the fact, rather than decided in advance. We report such checks separately and never let them decide a result, because it is far too easy to find a pattern once you already know the answer.
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.