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

Order Did Not Help, Mixing Might

By Joshua Bolick, Awareness Software Group ~10 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. A common idea in training is the curriculum: start with easy examples and work up to hard ones. We tested it on a memory task, using the same amount of compute for every approach, against simply practising the hard version from the start.

What we found. Working up in order was no faster. Surprisingly, a random mix of the easier versions early on ended noticeably more accurate on the hard version, and more consistently. We did not predict that in advance, so we are re-testing it on fresh runs before claiming it.

Why it matters. The lesson: an ordered curriculum is not automatically better than the obvious alternative, and a surprise found after the fact needs its own test.

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, 24 training runs of 6000 steps. The design, the shuffled control and the kill test were committed (e1db39b) before any run; one throwaway was disclosed in the pilot.

Program v2 Bucket A, item A14. Decisive computation: analysis/lag_staircase.py. Output: analysis/lag_staircase.json. Post-hoc, changing no verdict: analysis/lag_staircase_posthoc.py -> analysis/lag_staircase_posthoc.json.

The question

A curriculum -- easier examples first -- is the classic answer to "which data, in which order". A14 tests one at matched compute: J8's width-48 GRU copying a random stream at lag 8, against a staircase that trains at lags 2, 4, 6 for 500 steps each before lag 8, and A3's control, a shuffled mixture of the same lags in random order for the same 1500 steps. Same batch, length and steps, so the same FLOPs. Rate 0.01 (below T22's critical rate at every lag), eight receivers, 6000 steps. Endpoint: the step at which lag-8 held-out accuracy reaches 0.5.

**Easy-to-hard ordering did not help; mixing easy tasks in early may lift the ceiling.** Three ways to spend the same compute on a hard memory task: practise it from the start, work up through easier versions in order, or practise easier versions in a random mix first. Each line averages eight training runs. *Working up in order was no faster. The random mix of easier versions ended clearly higher on the hard task. That was not the test we set in advance, so we are now testing it properly on new runs.*

Kill test, fixed before execution: staircase minus fixed has an interval including zero or above it. Anchor: the fixed arm reproduces J8's own loop exactly. It holds.

Result: the kill test fires

Arm	Reached `0.5`	Step to `0.5`
fixed lag `8`	`7/8`	`2315` `[1585, 3045]`
staircase	`8/8`	`2551` `[1936, 3166]`
shuffled mixture	`8/8`	`2242` `[2140, 2343]`

Paired difference	Steps
staircase minus fixed	`+275` `[-214, +764]`
staircase minus shuffled	`+309` `[-267, +885]`
shuffled minus fixed	`-57` `[-713, +599]`

The staircase does not bring lag-8 learning sooner -- if anything later -- and the ordering earns no credit.

What the final accuracies show (post-hoc)

The lag-8 task plateaus not far above 0.5 for this model, so the arms may differ more in where they end than in when they cross it. At step 6000:

Final lag-`8` accuracy	Mean	Paired against fixed
fixed	`0.645` `[0.543, 0.747]`	--
staircase	`0.671` `[0.609, 0.734]`	`+0.027` `[-0.084, +0.137]`
shuffled mixture	`0.790` `[0.762, 0.817]`	`+0.145` `[+0.029, +0.261]`

The shuffled mixture also ends +0.118 [+0.045, +0.191] above the staircase. Mixing easier lags in early, in random order, may raise how far the model gets; putting them in order does not. This is post-hoc on eight receivers, so A15 preregisters it on fresh seeds.

What stands

Kill test fires: at matched compute an easy-to-hard lag staircase does not reach lag-8 accuracy 0.5 sooner.
Post-hoc: a shuffled mixture of the easier lags ends higher on lag 8 (+0.145), and more consistently, than either the fixed stream or the staircase.

Limits

One task, one width, one rate, one staircase schedule. The lag-8 task is near this model's capacity limit.

CONFIRMED 2026-09-28 by A15. On eight fresh seeds, with the endpoint fixed in advance, the shuffled mixture ends +0.215 [+0.135, +0.296] higher on lag-8 accuracy at step 6000 than the fixed stream. 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.
curriculum: Training on easier examples first and harder ones later, like a school syllabus, rather than on everything at once.
GRU: Gated Recurrent Unit. A compact design for processing sequences one item at a time, with internal switches controlling what it keeps in memory.
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.
plateau: A stretch of training where the model is not visibly improving. Often, though not always, followed by a sudden jump.
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.
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.
width: How many internal numbers a model uses at each layer. The usual way we vary model size in these experiments.