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

It Was the Rate, Not the Width

By Joshua Bolick, Awareness Software Group ~9 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, our widest model paid a smaller price for being taught an unneeded skill, but it was also trained more slowly, so we could not tell which one mattered. This time we changed only the learning speed, on a model half as wide.

What we found. The slower speed alone produced the same smaller price. Making the model wider, at the same speed, changed nothing. So how much an unneeded skill slows the others depends on how hard training pushes, not on how big the model is.

Why it matters. In practice: if something unneeded in the training mix is slowing a model down, training on it less often or more gently helps; a bigger model does not, at least at these sizes.

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, 12 training runs of 3000 steps. The design and kill test were committed (ab40ea6) before any run.

Program v2 Bucket T, item T7. Decisive computation: analysis/competition_rate_untied.py. Output: analysis/competition_rate_untied.json. Reproduce with python analysis/competition_rate_untied.py (about an hour on a throttled laptop CPU); --reuse re-derives every endpoint from the saved series, and was checked to print the identical summary.

The question

T6 found that at width 512 and its tuned rate 0.002, an unneeded skill (G) costs composite C 25% of its learning time, against about 35% at widths 64-256 -- but those narrower widths ran at the house rate 0.005, so width and rate changed together. T7 changes only the rate: T4's width-256 comparison (O17's task and seeds, every skill taught against G never taught) at 0.002.

**Slowing the learning rate eases the competition; making the model wider does not.** The extra time one skill takes when a small model is also taught a skill nothing else needs, as a percentage of its time without it. Two widths and two learning rates, changed one at a time. Lowering the learning rate at the same width cuts the cost from about 36% to 27%. Doubling the width at the same rate barely moves it. The competition depends on how hard training pushes, not on how big the model is.

Kill test, fixed before execution: C's share at width 256 and rate 0.002 lies within width 512's interval, [-0.270, -0.234] -- the lower rate alone explains T6's easing.

Result: the kill test fires

Width	Rate	`C` share	`F` share	`E` share
`256`	`0.005` (T4)	`-0.359` `[-0.439, -0.279]`	`-0.352`	`-0.271`
`256`	`0.002`	`-0.268` `[-0.288, -0.248]`	`-0.220`	`-0.221`
`512`	`0.002` (T6)	`-0.252` `[-0.270, -0.234]`	`-0.207`	`-0.170`

At the same width, lowering the rate from 0.005 to 0.002 moves C's share from 36% to 27%, into width 512's interval. At the same rate, doubling the width from 256 to 512 moves it from 27% to 25%, within noise. T6's easing was the rate.

What the thread now says

Across O20-O24 and T1-T7, one account fits every measurement:

The cost an unneeded skill imposes does not fall with width anywhere from 64 to 512 (O21, T4, T7).
It scales with how hard training pulls: halving how often the skill is taught roughly halves its cost for C (O24), and lowering the learning rate lowers it (T7).

That is what arXiv 2608.03573 calls norm-limited interference, and it is not the size-dependent interference arXiv 2605.29548 proposes -- at this scale, on this task. The second paper's range (4M-4B parameters) is far beyond width 512, so this bounds rather than refutes it.

What stands

Kill test fires. At width 256, rate 0.002 gives C a share of -0.268 [-0.288, -0.248], inside width 512's interval; the width effect T6 appeared to show is the rate.
Competition between skills here is set by the training pull (dose and rate), not by width.
Practical reading: if an unneeded objective is slowing a small model, a lower rate or a lower sampling share reduces its cost; a wider model does not.

Limits

One task family, widths 64-512, two rates.
A lower rate also slows learning overall; the share is relative to each run's own control, which removes that, but the absolute times differ.

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.

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.
parameters: The adjustable numbers inside a model. Training is the process of setting them. Model size is usually quoted as a count of these.
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.