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

On a Transformer, No Speed-Up

By Joshua Bolick, Awareness Software Group ~9 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: 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. On a small transformer, our first test of the easy-sum start leaned slightly in its favour, but most runs had not finished. We ran it again with twice as many runs and twice as long.

What we found. The lean reversed. Ordinary training reached the target on 21 of 32 runs against 16 with the easy start, and got there sooner. The easy-sum speed-up we found in the recurrent model does not carry over to this transformer.

Why it matters. One plausible reason: a transformer can look at both remembered symbols directly, so it does not need to build the machinery the easy sums help a recurrent model build. Either way, the result belongs to one kind of model, and we now say so everywhere.

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, 64 training runs of 16000 steps on thirty-two fresh seeds. The design and the kill test were committed (03ea64a) before any run.

Program v2 Bucket A, item A55. Decisive computation: analysis/sum_transformer_longer.py. Output: analysis/sum_transformer_longer.json.

The question

A54 found, on the tiny transformer, the early mix of easier sums reaching 0.8 on more seeds than plain training (9 against 5 of sixteen) in 8000 steps, with an interval reaching zero and most runs censored. Given thirty-two seeds and twice the budget, does the speed-up appear?

**On a small transformer the easy-sum start gives no speed-up.** The same comparison on a one-layer transformer, with thirty-two runs and twice the earlier training length. Each line counts how many runs have reached 80% accuracy on the hard sum by each step. *Given enough runs and time, ordinary training gets there more often and sooner. The easy-sum speed-up found in the recurrent model does not carry over to this transformer.*

Design: A54's transformer (one layer, four heads, width 48), arms and settings; thirty-two fresh seeds 22001-22032, 16000 steps; endpoint steps to 0.8 (never = 16000). Kill test, fixed before execution: mixed minus plain includes zero or lies above it. Anchor, in code: A54's plain run on its first seed, re-run, reproduces its first 8000 steps -- held.

Results

Arm (thirty-two fresh seeds, transformer)	Median steps to `0.8`	Reached `0.8`	Reached `0.9`
easier sums mixed first	`14585`	`16`	`9`
plain, learning-rate warm-up	`7735`	`21`	`17`

Mixed minus plain: +1424.1 [-1141.5, +3989.6] steps; ratio of mean steps 1.159.

The kill test fires -- and the point estimate has turned over. With enough seeds and budget, plain training reaches the criterion on more seeds and sooner than the mixture. The interval reaches zero, so the mixture is not shown to be slower; it is shown not to be faster. A54's lean towards the mixture was censoring noise.

What it says

The modular-sum speed-up is a property of the recurrent network, not of the task alone. On the GRU it holds at two widths, on two targets, against a tuned learning-rate warm-up and under decay (A40-A53); on a one-layer transformer trained the same way it does not appear, and leans the other way. A plausible reading, not tested here: a transformer can attend to both positions directly, so the easy sums do not give it the two-read-and-combine machinery a recurrent network has to build (A49, A52); for the transformer they are simply different data. The outside-venue draft's limits now say the result is specific to the recurrent model tested.

What stands

A55: kill test fires. Transformer, steps to 0.8: mixed minus plain +1424.1 [-1141.5, +3989.6]; plain reaches 0.8 on 21 of 32 against 16.
No speed-up on the transformer. The modular-sum result is architecture-specific on the evidence here.

Limits

One transformer configuration, one rate, not tuned for the transformer beyond two throwaways.

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.

architecture: The blueprint of a model: what components it has and how they connect. Two models can be the same size and completely different architectures.
GRU: Gated Recurrent Unit. A compact design for processing sequences one item at a time, with internal switches controlling what it keeps in memory.
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.
recurrent: A design that reads a sequence one item at a time, carrying memory forward. The main alternative is attention, which looks at everything at once.
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.
transformer: The architecture behind most modern large language models. It uses attention to look at every part of the input at once, rather than reading in order.
width: How many internal numbers a model uses at each layer. The usual way we vary model size in these experiments.