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

The Head Start Survives Tuning

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: Which parts of a model actually matter? – The parts that move most are not the parts that matter, and no single component is required -- the model routes around every freeze. What a part is worth shows up only when you remove it.

In plain English

What we asked. Our strongest result is that a small model learns much sooner if it is handed part of another model that has already learned the task. Every test of it used a learning rate nobody had tuned, so we re-ran it at the tuned one.

What we found. It held up. The head start was 39% of the time to learn instead of 41%, and handing over the same part from an untrained model still slowed learning down, so the benefit really does come from what the other model had learned. Getting the trained part was worth about four times as much as tuning the learning rate.

Why it matters. A result that survives the check most likely to break it is worth more than one that was never checked. This one has now survived a second task and a tuned baseline.

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, six seeds, 60 training runs. The design, the graft rule and the kill test were committed (eeacfbb) before the tuned rate was known.

Program v2 Bucket R, item R19. Decisive computation: analysis/transfer_at_tuned_rate.py. Output: analysis/transfer_at_tuned_rate.json. Reproduce with python analysis/transfer_at_tuned_rate.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

P1 is this programme's strongest positive: a donor's converged recurrent weights, grafted into a receiver at step 30, remove 35.8 steps from its time to the transition, where an untrained donor's weights delay it. Every record in that thread trains at 0.005. O15 showed a different thread had been compensating for an untuned rate, and R18 found P1's rate untuned too. Does the head start survive at the tuned rate, and is a donor worth more than simply tuning the rate?

**A head start from a trained model survives a properly tuned learning rate.** A small model is handed part of another, already-trained model early in training. The first four bars compare that with being handed the same part from an untrained model, at the old learning rate and the tuned one. The last bar compares getting the trained part at the old rate with getting nothing but a tuned rate. The head start barely changes when the learning rate is tuned, and an untrained part still slows learning, so the benefit comes from what the other model learned. Being handed trained parts is worth about four times what tuning the rate is worth here.

The tuned rate was not chosen here. It is read from O16's committed output: 0.006, the fastest rate on this substrate that loses no run to the gate and keeps final accuracy within 0.01 of the 0.002 control's.

Arms, at P1's 0.005 and at 0.006, every run P1's own run(): the donor (400 steps at the same rate), control, recurrent (P1's graft) and recurrent-untrained (P1's untrained-donor control).

The graft step, fixed before the tuned rate was known. A faster rate brings the transition sooner, so a graft at step 30 leaves less room before it and would shrink the head start by construction. The graft moves proportionally: 30 times the ratio of the two rates' mean control transitions, rounded. That gives step 27. P1's own step 30 is also run and reported.

Kill test: at 0.006 with the proportional graft, the recurrent head start has an interval including zero.

Anchor, in code: at 0.005, every seed's control and recurrent transitions equal P1's committed values. All six do.

Result: the kill test does not fire

Change in the transition against each seed's own control (steps; negative is sooner), P1's transition definition (floor-to-ceiling midpoint on a 5-step grid, not interpolated):

	`0.005` (P1, anchor)	`0.006` (tuned)
control transition	`86.7`	`79.2`
`recurrent` (converged donor)	`-35.83` `[-41.97, -29.70]`	`-30.83` `[-34.78, -26.88]`
`recurrent`, graft at P1's step `30`		`-29.17` `[-33.12, -25.22]`
`recurrent-untrained`	`+24.17` `[+19.01, +29.33]`	`+21.67` `[+18.96, +24.38]`
head start as a share of the control's time	`41.3%`	`38.9%`

The head start is six evaluation intervals clear of zero. As a share of the run it barely moves (41% to 39%), and the untrained-donor control still delays by about as much as before, so the contrast that makes P1 a statement about learned structure is intact.

The practical comparison. A receiver given a donor at the old rate reaches the transition 28.3 steps [-33.8, -22.9] sooner than a receiver alone at the tuned rate. Tuning the rate buys 7.5 steps here; the donor buys about four times that, at either rate.

What stands

P1 survives a tuned learning rate: -30.83 [-34.78, -26.88] at 0.006, 39% of the run.
Its controls survive with it: an untrained donor still delays (+21.67).
A donor is worth more than tuning the rate on this substrate, by about four to one.
The rest of the transfer thread (P6, P7, P8) shares P1's substrate and rate and was not re-run; R18's headroom (8.5 steps) is small beside their sizes, but only P1 has been tested.

Limits

One substrate, the one R18 could audit. The dispatch-copy claims (Q4, R12) remain UNTESTED.
The house rate was already close to tuned (8.5 steps of headroom), so this is a weaker test than O15's, where the rate sat 77 steps away. It shows P1 was not an artefact of the rate; it could not have shown a large effect vanishing, because the rate had little to give.
The donor's own economics (P8) were not re-priced: a donor trained at 0.006 is cheaper to make than one trained at 0.005, which would lower P8's break-even, but that is not measured here.

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.
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.
recurrent: A design that reads a sequence one item at a time, carrying memory forward. The main alternative is attention, which looks at everything at once.
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.

In plain English

The question

Result: the kill test does not fire

What stands

Limits

Terms on this page

Read next

Get new results as we publish them

Want this measured on your data?