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

The Level Was Not the Problem

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 makes training cheaper? – One thing has worked: stopping part of the training early saved about 7% with no loss of quality. Everything else tested has been matched by a simpler or cheaper method -- and in two cases the clever method was only winning because it was quietly being given more.

In plain English

What we asked. Last time a model's loss curve predicted its sudden jump in accuracy well when the vocabulary stayed the same, and badly when it changed. We guessed the reason: the loss starts from a different height for each vocabulary. So we measured it against a level set by the vocabulary instead.

What we found. It made no difference. The prediction stayed just as poor across vocabularies. With a large vocabulary the loss falls a long way before accuracy jumps; with a small one accuracy jumps first. They are simply measuring different things.

Why it matters. So our explanation was wrong, and the simple formula stays the better free way to predict the jump across different tasks.

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. No training: N20's committed curves, re-scored. The script and kill test were committed before it was run.

Program v2 Bucket N, item N25. Decisive computation: analysis/loss_vocab_normalised.py. Output: analysis/loss_vocab_normalised.json. Reproduce with python analysis/loss_vocab_normalised.py (seconds).

The question

N20 found that held-out loss crossing the midpoint of its own fall predicts when accuracy crosses 0.5 to 4.8 steps within vocabulary 32 but 15.7 across N6's grid, worse than N14's formula (9.6), and read the failure as the loss curve starting at the log of the vocabulary. The standing rule prefers a level fixed by the task over one normalised by the run. N25 fixes it: the loss crossing is the first evaluation at which held-out loss falls to half the natural log of the vocabulary size, with N20's leave-one-cell-out lead and N14's scoring unchanged.

**Setting the loss level by the task does not fix the signal across vocabularies.** The loss curve used to predict when accuracy jumps, with its crossing point defined two ways: the midpoint of each run's own fall (first bar of each pair), or a level set by the vocabulary size (second bar). The formula it competes with is off by about 10. Both definitions work well with 32 symbols and badly with 8 or 128. The problem is not where the level is drawn: across vocabularies, loss and accuracy measure different things, so a formula remains the better free predictor.

Kill test, fixed before execution: the error still exceeds 9.6 steps.

Result: the kill test fires

Loss crossing	Cell error, all 17 cells	Within vocabulary `32`
N20: midpoint of the run's own fall	`15.69` `[3.15, 28.22]`	`4.8`
N25: half of `ln V`, fixed by the task	`15.83` `[4.16, 27.49]`	`5.97`

The vocabulary columns that broke N20 break this too: the lead from loss crossing to accuracy crossing is +104.7 steps at vocabulary 128 (N20: +107.7) and -6 to -12 at vocabulary 8 (N20: -12 to -15). Moving the loss level to one set by the vocabulary changes almost nothing.

What this says

N20's explanation -- that the loss curve's starting height moved the midpoint -- is not what breaks the loss signal. With the level fixed by the task the leads barely move, so across vocabularies held-out loss and held-out accuracy are tracking different things: at vocabulary 128 the loss falls a long way (the model improving its probabilities over a large vocabulary) long before argmax accuracy reaches 0.5, and at vocabulary 8 accuracy reaches 0.5 before the loss has fallen far. A free instrument calibrated on one vocabulary does not transfer to another, and the offline formula remains the best free instrument across N6's grid.

What stands

Kill test fires. A task-fixed loss level scores 15.83 steps, no better than N20's 15.69.
The loss signal's vocabulary dependence is not a normalisation artefact; the thread closes here with the formula ahead across vocabularies and the loss signal ahead within one.

Limits

One alternative level (0.5 ln V), fixed in advance; others were not tried, deliberately, to avoid searching levels until one fits.
N6's grid only.

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.
argmax: Short for 'the argument that maximises'. A model outputs a score for every possible next token; the argmax is simply whichever one scored highest. Accuracy is measured on the argmax, which is why accuracy is a thresholded quantity: only the winner counts, and by how much it won is discarded.
calibrated: A model is calibrated when its confidence matches how often it is right: answers it gives with 90% confidence should be right about 90% of the time.
held-out: Data the model was never trained on, kept back specifically to test it. Scoring a model on data it has already seen measures memorisation, not learning.
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.
normalisation: Dividing a measurement by something else to make it comparable across cases. What you divide by is a choice, and it can change the answer as much as the data does.
vocabulary: The set of distinct symbols a model can read and produce.