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

One Nearby Version Does Most of It

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. Our warm-up mixed three easier versions of a hard memory task. Was the mixture itself important, or would one easier version do?

What we found. One easier version did most of the work, as long as it was close to the real task. Remembering 6 steps back, before a task that needs 8, lifted the final result from 58 to 76 percent on every run; the full mixture added at most a little more. Remembering 4 steps back, further from the target, did not reliably help.

Why it matters. In practice: warm up on a slightly easier version of the exact thing you want learned. Variety for its own sake is not the ingredient.

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 6000 steps. The design and the kill test were committed (21d8935) before any run.

Program v2 Bucket A, item A25. Decisive computation: analysis/warmup_single.py. Output: analysis/warmup_single.json.

The question

A21 found the warm-up must be related to the target, and A17 that variety kept up for the whole run hurts. Every helpful warm-up so far mixed three easier lags, so does the variety within the warm-up matter, or would one easier version do?

One easier version close to the target does most of the work. Every model trains for the same 6,000 steps and ends on the hard task: remember a symbol 8 steps back. The first 1,500 steps are spent on nothing different, on a 4-step version only, on a 6-step version only, or on a random mix of 2, 4 and 6 steps. Error bars are 95% confidence intervals. The 6-step version alone gets most of the way (76% against 58%, on every run); the mix adds at most a little. The 4-step version alone does not reliably help. What matters is an easier version close to the target, not variety.

Design: A15's task, rate 0.01 and seeds (J8's donor seeds), 6000 steps. 1500-step warm-ups of lag 4 only, lag 6 only, or lags 2/4/6 at random (A15's arm), then lag 8; plus fixed lag 8. Kill test, fixed before execution: the mix minus the better single-lag arm includes zero or lies below it. Anchor, in code: fixed and mix reproduce A15 exactly -- held on all eight seeds.

Results

Arm	Final lag-`8` accuracy	Minus fixed, paired	Seeds above fixed
fixed	`0.575` `[0.492, 0.657]`	--	--
lag `4` only	`0.625` `[0.526, 0.724]`	`+0.051` `[-0.082, +0.183]`	`5` of `8`
lag `6` only	`0.761` `[0.748, 0.775]`	`+0.187` `[+0.096, +0.278]`	`8` of `8`
mix `2`/`4`/`6`	`0.790` `[0.765, 0.815]`	`+0.215` `[+0.135, +0.296]`	`8` of `8`

The kill test fires. The mix minus lag 6 alone is +0.029 [-0.003, +0.060]: the mix is higher on 7 of 8 seeds, but the interval reaches zero, so the extra from variety is undecided and at most small. A single easier version -- the nearest one tried -- gives most of the gain. Lag 6 alone is also the most consistent arm in the programme: every seed ends between 0.74 and 0.79.

Which easier version matters. Lag 4 alone does not reliably help (+0.051, interval including zero, one seed at 0.89 and the rest between 0.52 and 0.65). The lag closer to the target does; the one further away does little.

What it says

The recipe narrows again. With A21 and A23: a short phase of an easier version close to the target does most of what the three-lag mix does. That the mix still matches lag 6 alone when it spends only a third of its warm-up on lag 6 -- and that A23's 100-step mix, about 33 steps of lag 6, gives the whole gain -- leaves two readings open: the mix's other lags contribute something, or very little lag-6 exposure is needed. Both are cheap to separate. A27 asks whether nearness is the rule (lags 5, 6 and 7 alone).

What stands

A25: kill test fires. The mix beats the better single lag by +0.029 [-0.003, +0.060]: undecided, at most small.
Lag 6 alone gives +0.187 [+0.096, +0.278], on every seed. Lag 4 alone does not reliably help.

Limits

Two single lags; whether nearness is the rule, and where it stops (lag 7 is nearly the target), is A27.
One task, one width, one warm-up length (1500, longer than A23 shows is needed).

CORRECTED 2026-09-28 by A27. The reading "nearness to the target" does not hold. With lags 5 and 7 added (same code and seeds), single-lag warm-ups before lag 8 peak at lag 5 (+0.246), and lag 7 helps less than lag 6 (+0.159). Lag 6 beating lag 4 was the right-hand side of a peak, not a trend toward the target. Text and numbers above unchanged.

QUALIFIED 2026-09-28 by A29: the baseline's learning rate was never tuned. This record ran at rate 0.01. At 0.005 the plain lag-8 run reaches 0.763 at step 6000 (against 0.575 at 0.01), and the 1500-step warm-up adds nothing (-0.034 [-0.104, +0.035]). The within-rate comparisons above stand as measured; any reading of them as an efficiency gain over the best plain alternative is suspended until A31 tunes both arms. 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.
baseline: The thing you compare against. A result without one is not a result.
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.
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.