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 Faster Rate, and Still Slower

By Joshua Bolick, Awareness Software Group ~9 min read August 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. The easy-start recipe slowed learning on a random lookup table. But we had only ever tuned the learning rate for ordinary training, never for the easy start. Maybe it was just set wrong.

What we found. A faster rate did help it, but not enough. At its best rate it was still about 1,600 steps slower than ordinary training at its best, and slower at every rate where the two could be compared directly.

Why it matters. So the slow-down is real, not a setting. On a task with no rule for the easier versions to share, starting easy costs time.

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 new training runs of 24000 steps plus 48 re-used from A66. The design and the kill test were committed (c2975fd) before any run.

Program v2 Bucket A, item A69. Decisive computation: analysis/table_mixed_rate.py. Output: analysis/table_mixed_rate.json.

The question

A66 found the easy warm-up slower than plain training on a random 20-symbol lookup table, by +3039.4 [+2328.3, +3750.4] steps against plain at its better peak (0.005) -- but the mixed arm ran at 0.003, and no mixed arm in this thread had ever been rate-tuned. Is the slow-down the mixed arm's rate?

**A faster learning rate helps the easy start, and it is still slower.** Both recipes were tried at several learning rates on the random lookup table. Lower is faster. Ordinary training was tried up to 0.005 and the easy-start recipe up to 0.008. *Both get faster at higher rates, but the easy start stays slower wherever the two can be compared, and its best is still about 1,600 steps behind ordinary training's best.*

Design: A66's loop, table, seeds and budget; the mixed arm at constant rates 0.005 and 0.008. The best of the three mixed rates against A66's best plain cell, paired. Kill test, fixed before execution: best mixed minus best plain lies wholly below zero (A66's slow-down was the rate). Anchor, in code: re-running A66's mixed 0.003 cell on its first seed reproduces it -- held.

Results

Arm (sixteen seeds, random `20`-symbol table)	Mean steps to `0.9`	Median	Solved
easy start, `0.003` (A66)	`8124`	`8160`	`16`
easy start, `0.005` (new)	`7250`	`6920`	`16`
easy start, `0.008` (new)	`6683`	`6520`	`16`
plain, `0.003` (A66)	`6575`	`6375`	`16`
plain, `0.005` (A66)	`5084`	`4920`	`16`

Best mixed (0.008) minus best plain (0.005): +1598.8 [+995.1, +2202.4] steps.
At the same rate, 0.005: mixed minus plain +2165.6 [+1170.0, +3161.3] steps.

The kill test does not fire. A higher rate helps the easy start (8124 to 6683 mean steps), but at its best rate tried it is still about 1600 steps slower than plain at plain's best, and slower at every rate where the two can be compared directly (0.003: +1548.8; 0.005: +2165.6).

What it says

A66's slow-down is not the mixed arm's untuned rate. On a random table, practising the same entries on nearer symbols costs time at every rate tried. With A66, this supports the reading that the easy start helps when the easy versions share the hard task's rule -- and on a table with no rule, it does not.

Both arms' best rates are at the tops of their grids (mixed 0.008, plain 0.005), so neither optimum is found. Each arm improved with its rate, plain by 1491 steps from 0.003 to 0.005 and mixed by 1441 from 0.003 to 0.008; nothing here says the mixed arm would overtake plain at still higher rates, and nothing rules it out. A70 runs both at 0.008 and 0.012.

The best mixed cell is picked from three on the same seeds the comparison uses, which favours the mixed arm, so the selection works against this result.

What stands

A69: kill test does not fire. Best mixed (0.008) minus best plain (0.005) +1598.8 [+995.1, +2202.4] steps; slower at every shared rate.

Limits

Both best rates at grid tops (A70). One table, one size, width 48, sixteen seeds.

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.

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.
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.