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

Half the Rate Matches the Warm-Up

By Joshua Bolick, Awareness Software Group ~11 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. For a day we reported that warming a small model up on easier versions of a task made it much better at the hard version. Every one of those comparisons trained the plain model at one setting of the learning rate, the size of each training step, which we had never tuned for this task.

What we found. At half that learning rate, the plain model did about as well as the warmed-up one had (76 percent against 79), and the warm-up then added nothing at all. The warm-up had been rescuing training from a poorly chosen setting, not beating the best ordinary training.

Why it matters. We have marked eleven earlier pages with this, put our outside write-up on hold, and the next experiments tune both versions properly. The lesson, which this project had already learned once: tune the ordinary method before crediting a new one.

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, 16 training runs of 6000 steps. The design and the kill test were committed (8580b61) before any run.

Program v2 Bucket A, item A29. Decisive computation: analysis/warmup_rate.py. Output: analysis/warmup_rate.json.

The question

A19 and A24 found the varied warm-up gives a steady +0.05 at width 96, against +0.215 at width 48. To keep width 96 learning, A19 also halved the rate from 0.01 to 0.005. Was the smaller gain the width or the rate?

**Training more slowly does what the warm-up did.** The same small model and hard task, trained with and without the easy-task warm-up, at the learning rate every earlier warm-up experiment used (0.01) and at half of it (0.005). The learning rate sets how large each training step is. Error bars are 95% confidence intervals. At the usual rate the warm-up lifts the result from 58% to 79%. At half the rate the plain run reaches 76% on its own, and the warm-up adds nothing. The earlier gains were measured against a baseline trained at a poorly chosen rate.

Design: width 48, A15's task (lag 8, length 16) and seeds (J8's donor seeds), 6000 steps, rate 0.005; fixed lag 8 against A15's 1500-step shuffled warm-up of lags 2/4/6. Kill test, fixed before execution: shuffled minus fixed has an interval including zero or lying below +0.1.

Results

Final lag-`8` accuracy	Rate `0.01` (A15)	Rate `0.005` (A29)
fixed lag `8`	`0.575` `[0.492, 0.657]`	`0.763` `[0.715, 0.812]`
shuffled warm-up, then lag `8`	`0.790` `[0.765, 0.815]`	`0.729` `[0.696, 0.762]`
shuffled minus fixed, paired	`+0.215` `[+0.135, +0.296]`	`-0.034` `[-0.104, +0.035]`

The kill test fires, and the answer is larger than the question. At rate 0.005 the warm-up does not help at all (-0.034, 2 of 8 seeds higher). The reason is in the fixed row: halving the rate lifts the plain run from 0.575 to 0.763, nearly to where the warm-up had lifted it at 0.01 (0.790). The warm-up's gain at 0.01 is matched by simply training at half the rate, with no warm-up and no extra data.

What this changes

Every copy-task record in the data-mixture thread compared against an untuned baseline. A14, A15, A16, A17, A20, A21, A23, A25, A26 and A27 all ran at rate 0.01, which was never swept for this task, and at which the plain run is far from its best. This is the failure the standing rule names -- tune the baseline's learning rate before crediting a speed-up, and check the archive for the sweep first -- and the archive had the warning for a closely related substrate: O16 tuned the width-48 GRU on delayed copy at lag 4 to 0.006, and found rates of 0.008 and above learn sooner but end worse. A14 chose 0.01 without checking it. Each of those records carries a banner from today.

What survives is narrower and still interesting. The within-rate comparisons stand as measured: at rate 0.01, related (A21), brief (A23), early (A26) easier material at the right distance (A27) rescues training that the rate would otherwise send somewhere worse. That reads like a data warm-up standing in for a learning-rate warm-up -- an early phase that stops a too-high rate from committing the model to a poor solution -- and A26's critical window is what such a mechanism would look like. What does not survive, until tested at a tuned rate, is any claim that the warm-up makes training more efficient than the best plain alternative.

The second task is not yet cleared either way. A18's modular-sum result ran at 0.005, also unswept for that task.

Generated

A31 -- the copy task tuned: fixed and shuffled at rates 0.002, 0.003 and 0.007 (with 0.005 from A29 and 0.01 from A15). Kill test: the best shuffled arm minus the best fixed arm, each at its own best rate, includes zero or lies below it.
A32 -- the same for A18's modular-sum task.
A33 -- a learning-rate warm-up against the data warm-up at rate 0.01: is the data warm-up doing a rate warm-up's job?

What stands

A29: kill test fires. At rate 0.005 the warm-up gives -0.034 [-0.104, +0.035].
Halving the rate alone lifts the plain run from 0.575 to 0.763, close to the warm-up's 0.790 at 0.01.
The copy-task warm-up records compared against an untuned baseline; their efficiency reading is suspended until A31.

Limits

Two rates. Whether a tuned rate leaves any room for the warm-up is A31.

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.
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.
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.