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

Four Agree, One Took a Shortcut

By Joshua Bolick, Awareness Software Group ~11 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. Our last result suggested that what matters is the total amount of material a model sees per training step, however it is packaged. So we gave it the same amount five ways, from a few long examples to many very short ones.

What we found. Four of the five took almost exactly the same time. The fifth looked four times faster, but its examples were so short that 'half right' meant getting one easy position right, which the model learns almost at once. Looking closely also showed we had miscounted the scored positions in two earlier write-ups; we have corrected both, and no conclusion changes.

Why it matters. The lesson: check what a score can be earned by. If one easy part can carry it, you are measuring the easy part.

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, 40 training runs, plus two diagnostic reruns. The design and kill test were committed (d7762fe) before any run. Two design faults were found on reading the result and are reported here, not repaired.

Program v2 Bucket T, item T14. Decisive computation: analysis/content_per_step.py. Output: analysis/content_per_step.json. Post-hoc diagnosis, changing no verdict: analysis/content_per_step_positions.py -> analysis/content_per_step_positions.json.

The question

T13 found batch size and sequence length act as substitutes on a random-stream copy task. T14 asked whether scored content per step -- batch times scored positions per sequence -- sets the pace: five cells meant to carry 768 scored positions per step, as batch 16, 32, 64, 128, 256 at length 52, 28, 16, 10, 7. Kill test: the slowest cell's mean step exceeds the fastest's by more than 25%.

**Four ways of packing a training step agree; the fifth took a shortcut.** Five settings meant to give the model the same amount of material per training step, packed as a few long sequences or many short ones. The last setting's sequences are so short that only two positions are scored, both right at the start. *The first four take almost exactly the same time. The fifth looks four times faster, but only because 'half right' there means getting one easy position right.*

**The first copied position is learned long before the rest.** The task asks the model to repeat the symbol it saw four places earlier. Each line is one position in the sequence. Position 5 copies the first symbol, right after the sequence's opening marker. *Position 5 is solved within a few dozen steps; the others take five to six times longer. Any score that can be reached by solving that one position measures the shortcut, not the skill.*

Result: the kill test fires

Batch x length	Scored positions per step (actual)	Step to `0.5`
`16 x 52`	`752`	`163.9` `[158.7, 169.2]`
`32 x 28`	`736`	`165.5` `[157.1, 174.0]`
`64 x 16`	`704`	`166.3` `[157.3, 175.3]`
`128 x 10`	`640`	`159.4` `[152.8, 166.1]`
`256 x 7`	`512`	`38.6` `[35.2, 42.0]`

Slowest over fastest: 4.31, against a limit of 1.25.

Two design faults, found on reading the result

1. The task scores length - 5 positions, not length - 4. The announcement token occupies position 0, so the first target sits at position lag + 1 = 5. The cells were not equal: they carried 752 down to 512 scored positions per step. The same wrong count appears in T11's "exponent against scored positions" and in T13's text; both now carry correction banners.

2. At length 7, the endpoint measures a shortcut. Only positions 5 and 6 are scored, so held-out accuracy 0.5 is reached by solving one of the two. A diagnostic rerun of the first receiver, reproducing T14's series exactly, records accuracy per position:

Cell	Position `5` reaches `0.5`	Later positions reach `0.5`	Mean reaches `0.5`
`256 x 7`	step `22.9`	`6`: `55.5`	`41.9`
`128 x 10`	step `34.7`	`6`-`9`: `163.8`-`224.6`	`158.4`

Position 5 -- copying the first data token, right after the announcement -- is learned in a few dozen steps at either length; the general copy at later positions takes five to six times longer. At length 7, both scored positions sit next to the start, so the cell's step to 0.5 times the shortcut, not the copy.

What the four comparable cells show

Setting aside the shortcut cell, the four cells from 16 x 52 to 128 x 10 agree within 4.3% (159.4-166.3) while their scored content per step spans 640-752 and their batch spans eight-fold. That is what the substitute reading predicts, but it is four cells, read after a fault was found, and the content was not equal as designed: it is not a result. T16 reruns it with the correct count and a length floor that keeps the endpoint on the general copy.

What stands

Kill test fires, on a cell whose endpoint measured something else. The verdict stands as recorded.
A fixed accuracy level means different things at different lengths when some positions are much easier than others: at length 7, 0.5 is one easy position. Check which positions a level can be reached by.
Correction: scored positions on this task are length - 5. T11's positions exponent becomes -0.237 [-0.257, -0.217] (was -0.262); its length exponent, its kill test and T13's verdict are unaffected.

Limits

The easy-position effect also touches T11 and T13's shortest cells (length 8, three scored positions, one of them position 5). It makes short sequences look faster, so if anything it flattens the context effect those records report.

FOLLOWED UP 2026-09-27 by T16. Rerun with the correct count and every length at least 16: four cells of equal scored content per step agree within 2.4%, and a cell with four times the content is faster than all of them. Text and numbers above unchanged.

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.
batch size: How many examples the model looks at before updating itself once. Bigger batches give a steadier but more expensive update.
exponent: The number in a power law that says how strongly one quantity responds to another. A bigger exponent means a steeper response to the same doubling.
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.
post hoc: Worked out after the fact, rather than decided in advance. We report such checks separately and never let them decide a result, because it is far too easy to find a pattern once you already know the answer.