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 Better Warm-Up, but No Gain

By Joshua Bolick, Awareness Software Group ~10 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: 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 first text experiment picked a poor set of easy documents. We rebuilt it with easy text from 86 different files and chose the learning rate with a short test first.

What we found. Starting on easy text now got ahead of a gentle learning-rate warm-up early in training, and the two finished level. But plain training at a constant rate was ahead of both the whole way.

Why it matters. So on a model this small, neither starting trick is needed. The published result we are testing would matter where a constant rate struggles, so that is the next test.

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, 3 sweep runs and 48 training runs of 4000 steps on sixteen seeds. The design, the rate-sweep rule, the reported loss comparisons and the kill test were committed (ac7856c) before any run.

Program v2 Bucket A, item A74. Decisive computation: analysis/text_curriculum_varied.py. Output: analysis/text_curriculum_varied.json.

The question

A72 tested arXiv 2506.11300's curriculum (easy documents first, as a warm-up) against a plain learning-rate warm-up on a small character-level model and found a constant rate ahead of both -- but its "easy third" was four documents, 46% of it two duplicated generated indexes. With a varied easy pool and a tuned rate, does the curriculum beat a plain warm-up?

**The curriculum beats a warm-up early, and a constant rate beats both.** The same three ways of starting as before, with the easy documents now drawn from 86 different files and the learning rate chosen by a short sweep. *Starting on easy text gets ahead of a learning-rate warm-up early on, but a plain constant rate is ahead of both throughout. At this scale neither starting trick is needed.*

Design: A72's model and arms on the same pinned prose with generated files excluded (39111 chunks); the easy pool built from the easiest documents with no document above 5% of it -- 86 documents; the rate swept first by a fixed rule (constant arm, one seed, lowest final loss): 0.002 gave 1.3061, 0.003 gave 1.3049, 0.005 gave 1.3201, so 0.003, inside the grid. Three arms at 0.003, sixteen seeds: constant, curriculum (1000 easy steps), rate warm-up (1000 steps). Kill test, fixed before execution: curriculum minus rate warm-up in steps to the constant arm's final loss includes zero or lies above it. The paired loss differences were named as reported quantities in advance. Anchor, in code: the corpus hash -- held.

Results

Arm	Mean steps to target	Reached (of `16`)	Final held-out loss
constant	`3806`	`13`	`1.3033`
curriculum	`3931`	`4`	`1.3084`
rate warm-up	`3944`	`3`	`1.3075`

Paired difference	Steps to target	Loss at `1000`	Loss at `2000`	Mean loss over the run	Final loss
curriculum - warm-up	`-12.5` `[-112.1, +87.1]`	`-0.016`	`-0.009`	`-0.094`	`+0.001` `[-0.002, +0.004]`
curriculum - constant	`+125.0` `[+0.4, +249.6]`	`+0.058`	`+0.008`	`+0.018`	`+0.005`
warm-up - constant	`+137.5` `[+37.9, +237.1]`	`+0.074`	`+0.016`	`+0.112`	`+0.004`

The kill test fires. In steps to target the curriculum and the warm-up are indistinguishable.

What it says

With the easy pool fixed, the curriculum behaves like a better warm-up than the warm-up: it is ahead of it through the first half of training (its held-out loss lower by 0.016 at step 1000), because a warm-up spends its first steps at a tiny rate while the curriculum trains at full rate on easy text. By the end the two are the same. But neither beats the constant rate, which is ahead of both throughout and at the end. That is A72's answer again, now without its design fault.

So at this scale and rate the question the paper leaves open does not arise: a warm-up is not needed (the constant rate trains stably), and the curriculum's easy phase costs a little rather than helping. The paper's comparison -- a curriculum against a constant rate with no warm-up -- would matter where a constant rate struggles, which on this model means a higher rate. A78 tests exactly that.

What stands

A74: kill test fires. Curriculum minus warm-up -12.5 [-112.1, +87.1] steps; the curriculum leads the warm-up early in loss and ties it at the end; the constant rate beats both. A72's conclusion holds with a varied easy pool.

Limits

A one-layer GRU of width 128, 4000 steps, 3 MB of text: far from the paper's scale. One rate; one easy-pool rule.
The rate sweep was one seed (fixed in advance); the three swept rates were within 0.015 of each other in final loss.

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.

curriculum: Training on easier examples first and harder ones later, like a school syllabus, rather than on everything at once.
GRU: Gated Recurrent Unit. A compact design for processing sequences one item at a time, with internal switches controlling what it keeps in memory.
held-out: Data the model was never trained on, kept back specifically to test it. Scoring a model on data it has already seen measures memorisation, not learning.
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.