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

A Sweet Spot, Not the Nearest

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. We had found that warming a model up on one slightly easier version of a memory task helps it learn the hard version, and guessed that the closer the easy version is to the hard one, the better. We tested that guess.

What we found. It was wrong. Before a task that needs remembering 8 steps back, warming up on 5 steps back worked best, lifting the result from 58 to 82 percent on every run. Six and seven steps back helped less, and four steps back barely helped at all.

Why it matters. The most useful warm-up is easier by a clear margin, but not so easy that it can be solved some other way. We have corrected the earlier page that suggested nearer is better.

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 (6b4d460) before any run.

Program v2 Bucket A, item A27. Decisive computation: analysis/warmup_nearness.py. Output: analysis/warmup_nearness.json.

The question

A25 found a 1500-step warm-up on lag 6 alone gives most of the warm-up gain before a lag-8 target, and lag 4 alone does not reliably help. It read that as nearness to the target. Is nearness the rule? If so, lag 7 should do at least as well as lag 6, and lag 5 less.

**The best warm-up is easier by a margin, not the nearest.** Every model ends on the hard task, remembering a symbol 8 steps back, after the same 6,000 steps. The first 1,500 steps are spent on one easier version: remembering 4, 5, 6 or 7 steps back. Error bars are 95% confidence intervals. The 4-step bar comes from the previous experiment, with the same code and runs. Five steps back is best (82% against 58%, on every run). Nearer versions help less, and one much easier does not reliably help at all. The useful warm-up is easier by a margin, but not so easy that it is solved some other way.

Design: A25's code, task, rate 0.01 and seeds (J8's donor seeds), 6000 steps. 1500-step warm-ups of lag 5, 6 or 7 only, then lag 8; plus fixed. Kill test, fixed before execution: lag 7 minus lag 5 includes zero or lies below it. Anchor, in code: fixed and lag 6 reproduce A25 exactly -- held on all eight seeds.

Results

Warm-up lag (target `8`)	Final lag-`8` accuracy	Minus fixed, paired	Seeds above fixed
none (fixed)	`0.575` `[0.492, 0.657]`	--	--
`4` (A25, same code and seeds)	`0.625` `[0.526, 0.724]`	`+0.051` `[-0.082, +0.183]`	`5` of `8`
`5`	`0.821` `[0.789, 0.852]`	`+0.246` `[+0.161, +0.331]`	`8` of `8`
`6`	`0.761` `[0.748, 0.775]`	`+0.187` `[+0.096, +0.278]`	`8` of `8`
`7`	`0.734` `[0.702, 0.766]`	`+0.159` `[+0.064, +0.254]`	`8` of `8`

The kill test fires, the other way round from the prediction. Lag 7 minus lag 5 is -0.087 [-0.130, -0.044], lag 5 higher on every seed. Past lag 5, the nearer the warm-up is to the target, the less it helps; at lag 4 the benefit collapses. The curve peaks at lag 5 -- the best single-lag warm-up tried, as good as the three-lag mix (A25's 0.790) or A23's 100-step mix (0.817).

A25's reading, "nearness", was a two-point line through the right-hand side of a peak. It is corrected by banner.

What it says

An easier version helps most when it is easier by a margin, but not so easy it is solved another way. Two plausible stories, both testable: a lag close to the target is barely easier, so its warm-up is nearly the hard task itself (lag 7's warm-up reached 0.171 on lag 8 by step 1500, lag 5's only 0.041); and a lag too short can be solved by a shortcut that does not carry over. Where the peak sits relative to the target -- a fixed gap, a fixed ratio, or something set by what the model can already do -- is the question worth carrying forward.

A minimal recipe is now on the table: one easier version at the sweet spot, very briefly, at the start (A23, A26, A27). A30 tests it: 100 steps of lag 5 alone.

What stands

A27: kill test fires. Lag 5 beats lag 7 by 0.087 [0.044, 0.130], on every seed.
Single-lag warm-ups before lag 8 peak at lag 5 (+0.246); lags 6 and 7 help less, lag 4 unreliably.
A25's "nearness" reading is corrected.

Limits

One target (lag 8); where the peak sits for another target is not known.
1500-step warm-ups; A23 showed 100 steps of the mix suffice, and whether 100 of lag 5 does is A30.

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.