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

Slower Than Tuned, Not Hotter

By Joshua Bolick, Awareness Software Group ~8 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: Is the task we are studying actually hard? – Often it is not. A rule from 1990 with no parameters beats the trained model on the task most of these results were measured on, and what an intervention costs is set by the task's own structure.

In plain English

What we asked. After a too-large learning rate undid some of our results, we checked every other result. A few ran on tasks where nobody had found the best rate yet. We found it for one of them.

What we found. That task learns fastest at a learning rate of 0.010, twice the 0.005 the earlier result used. So the earlier result was run slowly, not too fast, and its conclusion does not change.

Why it matters. Audits are mostly reassurance. Doing them anyway is how you find the ones that are not.

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, 30 training runs of 600 steps. The design and the kill test were committed (c39b955) before any run.

Program v2 Bucket A, item A38. Decisive computation: analysis/wide_vocab_rate.py. Output: analysis/wide_vocab_rate.json.

The question

A37 found two surviving claims with no tuned rate on their own task. R12 is one: it ran at 0.005 on dispatch-copy-wide-vocab, never swept. Did it run hot?

**This task learns fastest at a rate twice the one it was run at.** A copying task with a wide vocabulary, used by one earlier result (orange: the rate it was run at). Each bar is the number of training steps six runs needed to reach the task's success criterion. The green bar is the best rate by the programme's rule. *The earlier result ran slower than this task's best rate, not faster, so it did not have the too-large-rate problem that undid the warm-up results. Its conclusion does not move.*

Design: O19's recipe, unchanged, on R12's own loop and seeds: whole-run rates 0.003-0.014, R12's criterion transition and final accuracy; the tuned rate is the fastest whose runs all reach the criterion and whose mean final accuracy is within 0.01 of the lowest rate's. Kill test, fixed before execution: the tuned rate is below 0.005. Anchor, in code: at 0.005 every control reproduces R12's committed control exactly -- held.

Results

Rate	Mean criterion transition (steps)	Final accuracy
`0.003`	`206.7`	`0.522`
`0.005` (R12)	`153.3`	`0.548`
`0.007`	`133.3`	`0.558`
`0.010`	`123.3`	`0.562`
`0.014`	`130.0`	`0.563`

The kill test does not fire. The tuned rate is 0.010 -- the same as O19 found for the narrower dispatch copy -- so R12 ran below its task's tuned rate, not above it. Every run reaches the criterion at every rate.

What that means for R12. R12 is exposed to R18's opposite concern (a slow rate) rather than A37's: at 0.010 its control reaches the criterion about 30 steps sooner. R12's question was which of two predictors -- 1.40 times the control's transition, or rung index 4 -- locates the donor-length optimum. At the tuned rate the multiple predicts 1.40 x 123.3 = 172.6, whose nearest rung is still 200, the rung R12 measured. The rung the multiple points to does not move; whether the optimum itself moves at the faster rate is not tested here.

What stands

A38: kill test does not fire. dispatch-copy-wide-vocab tunes to 0.010; R12's 0.005 was not hot.
One claim (R2) now has no tuned rate on its own task -- its pool of several tasks has no single rate to sweep.

Limits

Six seeds and O19's rule, which trades speed against a 0.01 accuracy tolerance.

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.

accuracy: The fraction of answers a model gets right on questions it was not trained on.
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.

In plain English

The question

Results

What stands

Limits

Terms on this page

Read next

Get new results as we publish them

Want this measured on your data?