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

Less Competition at Width 512

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. Our widest models could not learn properly at our usual learning speed, so first we found the right speed for them. That confirmed our guess: the usual speed was too fast for a model that wide, and a slower one let it learn nearly four times sooner.

What we found. At the right speed, the widest models paid a smaller price for being taught an unneeded skill: about a quarter of their learning time instead of the third that every smaller size paid. That fits a 2026 paper's claim that bigger models suffer less interference.

Why it matters. But two things changed at once, the width and the learning speed, and a slower speed could shrink the interference on its own. So we are not claiming the width did it yet; the next test changes only the speed.

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, 24 training runs of 3000 steps at width 512. The design, tuning rule and kill test were committed (204775e) before any run.

Program v2 Bucket T, item T6. Decisive computation: analysis/competition_wide_tuned.py. Output: analysis/competition_wide_tuned.json. Reproduce with python analysis/competition_wide_tuned.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

T4 found the share of each composite skill's learning time that an unneeded skill (G) costs is flat from width 64 to 256 (C about 35%), and that at width 512 the house rate 0.005 made learning slower and unreliable. T6 tunes the rate at width 512, then repeats the comparison.

**At the widest size, with its rate tuned, an unneeded skill costs less.** How much longer one skill takes to learn when the model is also taught a skill nothing else needs, at four model widths. The three narrower widths used our usual learning rate; the widest needed a slower one to learn reliably, found first on separate runs. *At the widest size the cost drops from about a third to a quarter. But the width and the learning rate changed together, so either could be the reason; the next test changes only one of them.*

Part A: width 512, three throwaway seeds, rates 0.001-0.005; the tuned rate is the one, among those at which every seed acquires C, E and F, with the lowest mean acquisition of C. Part B: O20's paired comparison (every skill taught against G never taught) on O17's six seeds at the tuned rate, the share by O21's function.

Kill test, fixed before execution: C's share at width 512 overlaps T4's flat width-256 mean (-0.359). Precondition: C acquired on at least four seeds in both arms. It passes (six of six).

Part A: the house rate was the problem at width 512

Rate	Mean step `C` is acquired	Every seed acquires `C`, `E`, `F`
`0.001`	`607`	yes
`0.002`	`490`	yes
`0.003`	`620`	yes
`0.005` (house)	`1813`	no

T4's inference is confirmed: at width 512 the house rate is too high, and 0.002 learns C nearly four times as fast.

Part B: the kill test does not fire

Skill	Share at width `512`, rate `0.002`	Widths `64`-`256`, rate `0.005` (T4)
`C`	`-0.252` `[-0.270, -0.234]`	about `-0.35`
`F`	`-0.207` `[-0.245, -0.168]`	`-0.26` to `-0.37`
`E`	`-0.170` `[-0.234, -0.107]`	about `-0.27`

At width 512 with a tuned rate, an unneeded skill costs each composite less of its learning time -- a quarter for C rather than a third -- and C's interval excludes the flat value of the narrower widths. That is the first result here consistent with arXiv 2605.29548's claim that interference falls as models grow.

The confound, stated before anything is built on it

Width and learning rate changed together. Widths 64-256 ran at 0.005; width 512 ran at 0.002. A lower rate means smaller steps, and O24 and 2608.03573 both say this kind of interference scales with the gradient's pull -- so a lower rate could shrink the share by itself, at any width. This is the tied-grid fault this programme has recorded before (Q8): the design cannot tell "wider models compete less" from "a slower rate competes less". Generates T7, which untangles them by running width 256 at 0.002.

What stands

Kill test does not fire. At width 512 and rate 0.002, C's share is -0.252 [-0.270, -0.234], against about -0.35 at widths 64-256 and rate 0.005.
The house rate fails at width 512; 0.002 is its tuned rate on this task.
Width and rate are confounded here; which one eased the competition is T7's question.

Limits

Tuned on three throwaway seeds, one task, one tuning criterion.
The confound above means this record does not support "larger models compete less" on its own.

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.
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.
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.