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

Boost Then Decay

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: What actually makes training cheaper? – One thing has worked: stopping part of the training early saved about 7% with no loss of quality. Everything else tested has been matched by a simpler or cheaper method -- and in two cases the clever method was only winning because it was quietly being given more.

In plain English

What we asked. Our short, early burst of faster learning made a small model learn sooner. The standard schedule used for large models, which slows learning down near the end, made it finish more accurately. The two work at opposite ends of training, so we tried using both.

What we found. The combination learned as soon as the burst alone and finished about as accurately as the standard schedule: faster and more accurate than simply training at one steady setting.

Why it matters. Two honest caveats. The speed part was almost certain, because the slowing-down happens long after the model has learned; the accuracy part was the real test. And the burst itself has already failed to help on two other setups, so this is a result about one setup, not a recipe.

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. The design and kill test were committed (fb04b07) before any run.

Program v2 Bucket T, item T3. Decisive computation: analysis/boost_plus_decay.py. Output: analysis/boost_plus_decay.json. Reproduce with python analysis/boost_plus_decay.py (about eight minutes on a throttled laptop CPU); --reuse re-derives every endpoint from the saved runs, and was checked to print the identical summary.

The question

T2 found that on J8's substrate (width-48 GRU, delayed copy, lag 4) at the tuned rate 0.006, O18's short early boost reaches the transition fastest while warmup-stable-decay ends the most accurate, and that the two act at opposite ends of the run. T3 joins them: the 1.5x boost for steps 36-75, then linear decay to 0 over the last 20%, as warmup-stable-decay decays.

**Boost early, slow down late: one schedule gets both speed and accuracy.** Eight small models trained three ways: our short early boost alone, the standard warm-up-and-decay schedule, and the two combined (boost early, decay at the end). Left: steps until the models learn. Right: wrong answers per thousand at the end. The combined schedule learns as soon as the boost and finishes with about as few mistakes as warm-up-and-decay. On this setup it beats both of its parts, though the boost half has already been shown not to carry over to other setups.

Every run is J8's own train() through T2's optimiser wrapper, on J8's eight receivers.

Kill test, fixed before execution: the combined schedule is not both within one evaluation interval (5 steps) of the boost's transition and at least as accurate as warmup-stable-decay (its paired accuracy difference lying wholly below zero).

Anchors, in code: the boost and wsd arms re-run here equal T2's committed runs on every receiver. They do.

Result: the kill test does not fire

Arm	Transition (steps)	Final accuracy
`boost`	`66.99` `[64.09, 69.88]`	`0.9778` `[0.9756, 0.9801]`
`wsd`	`75.71` `[72.82, 78.60]`	`0.9877` `[0.9855, 0.9900]`
`boost+decay`	`67.39` `[64.49, 70.29]`	`0.9866` `[0.9838, 0.9894]`

Paired	Difference
`boost+decay` minus `boost`, transition	`+0.41` `[+0.32, +0.50]`
`boost+decay` minus `wsd`, final accuracy	`-0.0011` `[-0.0035, +0.0012]`

The combined schedule reaches the transition with the boost and ends as accurate as warmup-stable-decay. Against the tuned constant rate (T2's constant: 76.32 steps, 0.9793), it is about 9 steps sooner and 0.0073 more accurate.

What the test could and could not have shown

The speed half was nearly guaranteed by the design. The decay begins at step 368, some 300 steps after the transition, so it can barely move the transition: the +0.41 is the whole effect. The accuracy half was the real test, and it could have failed -- a model already past its transition at a boosted rate might not recover warmup-stable-decay's accuracy from a decay alone. It did.

What stands

Kill test does not fire. Boost-then-decay matches the boost's transition (+0.41, inside one evaluation interval) and warmup-stable-decay's final accuracy (-0.0011 [-0.0035, +0.0012]).
On this substrate, a two-part schedule beats both its parts: about 12% sooner than the tuned constant rate and more accurate than it.
It rests on O18's boost, which O19 found does not travel to a second task or a wider model; the decay half is a standard practice that likely does. Nothing here claims more than J8's substrate.

Limits

One substrate, eight receivers, 460 steps, one decay share and one boost; by O19 the boost part is substrate-specific.
The speed comparison was structurally easy, as described above.

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.
evaluation interval: The gap between successive checks of a model during training. It is the smallest difference in timing a measurement can resolve, so two results closer together than one interval cannot be told apart.
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.
width: How many internal numbers a model uses at each layer. The usual way we vary model size in these experiments.