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

Flat Through Width 256

By Joshua Bolick, Awareness Software Group ~10 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. Teaching a small model a skill it does not need slows its other skills. A 2026 paper argues this kind of interference shrinks as models get bigger, so we tried models up to eight times wider than our usual size.

What we found. Up to four times wider, nothing changed: the unneeded skill cost the others the same share of their learning time. At eight times wider, training itself went wrong at our usual learning rate: the biggest models learned more slowly than smaller ones, and some did not learn at all.

Why it matters. So on this task, a bigger model does not end the competition between skills, and the widest models need their learning speed tuned before they can tell us anything. That is the next step.

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, 25 training runs of 3000 steps at widths 256 and 512. The design and kill test were committed (a45b4c0) before any run, with the run time measured first.

Program v2 Bucket T, item T4. Decisive computation: analysis/competition_wide.py. Output: analysis/competition_wide.json. Reproduce with python analysis/competition_wide.py (about four hours on a throttled laptop CPU); --reuse re-derives every endpoint from the saved series, and was checked to print the identical summary.

The question

T1 read arXiv 2605.29548, which argues larger models learn more because interference between tasks falls with size. O21 found the share of a composite's time that an unneeded skill (G) costs barely moved between widths 64 and 128. T4 repeats O21's exact comparison at widths 256 and 512 -- O17's task and seeds, every skill taught against G never taught, O12's own loop with only the width changed, the share computed by O21's own function.

A model four times as wide pays the same share for an unneeded skill. How much longer two skills take to learn when a small model is also taught a skill nothing else needs, as a percentage of their time without it, at three model widths. The widest setting tried, 512, broke down at our usual learning rate and is not shown. Across a fourfold range of width the cost stays the same: about a third of the time for one skill and a quarter for the other. Making the model bigger did not make the skills stop competing on this task.

Kill test, fixed before execution: at width 512, F's share has an interval overlapping width 64's mean (-0.370). Precondition: F acquired on at least four seeds in both arms at each width -- it passes, narrowly (four at width 512). Anchor: O21's, reproduced exactly.

Result: the kill test fires -- on an interval that says nothing, from a width that broke

Width	`C` share	`F` share	`E` share
`64`	`-0.324` `[-0.766, +0.118]`	`-0.370` `[-0.622, -0.117]`	`-0.277` `[-0.470, -0.084]`
`128`	`-0.353` `[-0.399, -0.308]`	`-0.261` `[-0.333, -0.189]`	`-0.278` `[-0.339, -0.216]`
`256`	`-0.359` `[-0.439, -0.279]`	`-0.352` `[-0.443, -0.262]`	`-0.271` `[-0.384, -0.158]`
`512`	--	`-0.177` `[-0.917, +0.563]`	--

Through width 256 the share does not fall. An unneeded skill costs C about 35% of its time at every width from 64 to 256, and E about 27%; F wanders between 26% and 37% with no trend. Across a fourfold range of width, competition between skills does not ease on this task.

Width 512 cannot be read. At the house learning rate 0.005, width-512 models learn more slowly than width-256 ones -- C at 1902 steps against 755, A at 975 against 472 -- and unreliably: C is acquired on four of six control seeds and two of six G-never seeds, E on one, and one G-never run never learns even the look-ups. The kill test fires, but on an interval of [-0.92, +0.56] from four paired seeds, which excludes nothing. The pattern is the one this programme has now found three times (O15, R18, O19): a rate that suits one width is not the rate for another. That cause is inferred, not tested.

What stands

Across widths 64-256, the share of each composite's time that an unneeded skill costs is flat (C about 35%, E about 27%).
2605.29548's falling interference does not show over a fourfold width range on this task; whether it appears at larger widths is untested, because width 512 broke.
At width 512 the house rate fails: learning slows and becomes unreliable. Generates T6: tune the rate at width 512 by O16's rule, then repeat.

Limits

Six seeds per width; width 64's intervals are wide (O20).
A fixed task across widths makes the larger models progressively oversized (P11) -- here that is the question, not a confound, but any easing would be about capacity relative to this task.
Width 512's failure is attributed to the learning rate by analogy, not by test.

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.

In plain English

The question

Result: the kill test fires -- on an interval that says nothing, from a width that broke

What stands

Limits

Terms on this page

Read next

Get new results as we publish them

Want this measured on your data?