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

Fast or Accurate

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. Earlier we found a short, early burst of faster learning that helped one small model learn sooner. We had never compared it with the learning schedules people actually use to train large models, which our new habit of checking recent research turned up.

What we found. Our burst still got the models to the answer first. But the standard schedule, which ramps up, holds, and then slows learning down at the end, left them making the fewest mistakes by the end of training. Neither was best at both.

Why it matters. The two act at opposite ends of training, one early and one late, so the obvious next step is to use both, which is what we test next.

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. The design and kill test were committed (e9ffbc2) before any run, with the standard arms specified before execution.

Program v2 Bucket T, item T2 -- the first item generated by the standing source search (docs/sources.md). Decisive computation: analysis/boost_versus_schedules.py. Output: analysis/boost_versus_schedules.json. Reproduce with python analysis/boost_versus_schedules.py (about ten 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

O18 found that on J8's substrate (width-48 GRU, delayed copy, lag 4) at the tuned rate 0.006, a 1.5x learning-rate boost for steps 36-75 reaches the transition as soon as a whole-run 0.008 and keeps 0.006's accuracy; O19 found it does not travel. The first standing source search turned up the schedules the literature actually uses -- warm-up theory (arXiv 2510.03164) and warmup-stable-decay (PMLR v336) -- and O18 had never been priced against them.

**The short boost learns fastest; the standard warm-up-and-decay ends most accurate.** Eight small models trained with four learning-rate schedules: steady; our short early boost; the common warm-up, which starts from zero; and warm-up then decay, a schedule widely used for large models. Left: steps until the models learn. Right: wrong answers per thousand at the end. *Our boost gets there first, and the standard warm-up-and-decay finishes with the fewest mistakes. Neither wins both, and they act at opposite ends of training, so the next test combines them.*

Every run is J8's own train() on J8's eight receivers, schedules applied by wrapping the optimiser: constant (0.006), boost (O18's), warmup (linear 0 to 0.006 over the first 40 steps, then constant) and wsd (linear 0 to 0.009 over 40 steps, stable, linear decay to 0 over the last 20%).

Kill test, fixed before execution: a standard arm matches or beats the boost on both time to transition and final accuracy.

Anchors, in code: constant equals O16's committed runs and boost equals O18's, on every receiver. Both do.

Result: the kill test does not fire -- the schedules trade

Arm	Transition (steps)	Final accuracy
`constant`	`76.32` `[72.68, 79.95]`	`0.9793` `[0.9760, 0.9826]`
`boost`	`66.99` `[64.09, 69.88]`	`0.9778` `[0.9756, 0.9801]`
`warmup`	`89.99` `[87.23, 92.75]`	`0.9821` `[0.9782, 0.9860]`
`wsd`	`75.71` `[72.82, 78.60]`	`0.9877` `[0.9855, 0.9900]`

Paired against `boost`	Transition	Final accuracy
`warmup`	`+23.00` `[+21.99, +24.02]`	`+0.0042` `[+0.0006, +0.0079]`
`wsd`	`+8.72` `[+7.42, +10.03]`	`+0.0099` `[+0.0074, +0.0123]`

No standard schedule matches the boost on both endpoints, so the kill test does not fire. But neither does the boost win outright:

The boost is the fastest schedule here -- 8.7 steps ahead of warmup-stable-decay and 23 ahead of the common warm-up, which delays the transition by starting from zero.
Warmup-stable-decay ends the most accurate, 0.0099 above the boost and above every other arm, and the source says why: the decay at the end removes the noise a high stable rate carries.

The two schedules act at opposite ends of the run -- the boost early, the decay late -- so they are not rivals so much as candidate halves of one schedule.

What stands

Kill test does not fire. Warm-up is slower and only slightly more accurate; warmup-stable-decay is 8.7 steps slower and 0.0099 more accurate.
On this substrate the boost buys speed and warmup-stable-decay buys final accuracy.
Generates T3: the boost followed by warmup-stable-decay's closing decay, against both, at matched steps.

Limits

J8's substrate only, where O19 already showed the boost does not travel; 460 steps; one decay share (20%) and one warmup-stable-decay peak (0.009).
Time to transition is this programme's endpoint; a practitioner optimising final accuracy alone would read the table the other way round.

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.
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.
width: How many internal numbers a model uses at each layer. The usual way we vary model size in these experiments.

In plain English

The question

Result: the kill test does not fire -- the schedules trade

What stands

Limits

Terms on this page

Read next

Get new results as we publish them

Want this measured on your data?