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, Not Established

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: 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. Every result so far on the adding task used one kind of model, a small recurrent network. We tried the same comparison on a small transformer, the design behind modern language models.

What we found. The easy-sum start got more runs to the target, 9 of 16 against 5, and sooner on average. But most runs of both recipes had not reached the target by the end of training, so the difference is within noise.

Why it matters. Not a failure and not a success: undecided. The next test runs twice as long with twice the runs.

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, 32 training runs of 8000 steps on sixteen fresh seeds. The design, two disclosed throwaways and the kill test were committed (4d08178) before any run.

Program v2 Bucket A, item A54. Decisive computation: analysis/sum_mixture_transformer.py. Output: analysis/sum_mixture_transformer.json.

The question

Every modular-sum result so far is a GRU: the early mix of easier sums reaches the solution in about 40% fewer steps than the best plain recipe (A44). Does it hold on the programme's tiny transformer?

**On a small transformer, the easy-sum lead is not yet clear.** The same comparison on a different kind of model, a one-layer transformer. Each line counts how many of sixteen runs have reached 80% accuracy on the hard sum by each step. *The mix gets more runs there (9 against 5), but most runs of both recipes had not reached 80% by the end, so the difference is within noise. A longer test with more runs follows.*

Design: one layer, four heads, width 48; target (2, 7) at 0.003, 8000 steps, sixteen fresh seeds. Mixed (2000 steps of (2,3)/(2,4)/(2,5)) against plain with a 500-step rate warm-up. Endpoint: steps to 0.8 -- the throwaways showed the transformer can plateau just under 0.9, so the GRU's 0.9 would partly measure the plateau. Kill test, fixed before execution: mixed minus plain includes zero or lies above it. Anchor, in code: with the GRU restored, the loop reproduces A40 -- held.

Results

Arm (sixteen fresh seeds)	Median steps to `0.8`	Reached `0.8`	Reached `0.9`
easier sums mixed first	`4790`	`9`	`5`
plain, learning-rate warm-up	`8000` (budget)	`5`	`3`

Mixed minus plain in steps to 0.8: -1312.5 [-3258.5, +633.5]; ratio of mean steps 0.804.

The kill test fires. The mixture leads on the point estimate and in how many seeds reach the criterion (9 against 5), and the interval reaches zero. On this transformer, at this budget and sample, the speed-up is not established.

What it says

The transformer was less reliable than its throwaway suggested: fewer than a third of plain runs and just over half of mixed runs reached 0.8 in 8000 steps, so most of the paired differences are between one arm's crossing and the other's budget. The comparison is mostly censoring, which is the situation A41 warned about. It is not evidence the effect is absent on transformers; it is not evidence it is present. A55 gives it the budget and seeds to decide.

What stands

A54: kill test fires. Mixed minus plain in steps to 0.8: -1312.5 [-3258.5, +633.5]; reached 0.8 on 9 against 5 of sixteen.
The modular-sum speed-up is established on the GRU only.

Limits

Sixteen seeds, 8000 steps -- short for this transformer, which censors most runs. One configuration, one rate.

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.

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.
plateau: A stretch of training where the model is not visibly improving. Often, though not always, followed by a sudden jump.
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.