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

Longer Sequences Did Not Help

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: 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. The same paper whose batch-size formula fit our model also says longer training examples should make a model learn sooner. We tested that half by making each example up to eight times longer.

What we found. It made almost no difference. But our task builds each example by repeating one short pattern, so a longer example only held more copies of the same thing. We caught that when reading the result, said so, and set up a fair version where longer means more new material.

Why it matters. The useful point: more text is not more information if it repeats itself. Count what is new in your data, not how long it is.

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, anchor and kill test were committed (9a5a233) before any run.

Program v2 Bucket T, item T9. Decisive computation: analysis/context_law.py. Output: analysis/context_law.json. Reproduce with python analysis/context_law.py (about twenty minutes on a throttled laptop CPU); --reuse re-derives every endpoint from the saved series.

The question

arXiv 2511.16893 fits the step at which induction heads form as B^-0.37 * C^-0.62: longer contexts should bring the transition sooner, more strongly than bigger batches. T8 found the batch term holds on J8's recurrent network at a fixed rate (-0.315). T9 tests the context term on the same substrate: batch 64, rate 0.006, sequence length 8, 16, 32, 64 with the copy lag held at 4, eight receivers, T8's endpoint (first step at held-out accuracy 0.5) and per-receiver log-log slope.

**Longer sequences of the same pattern did not speed learning.** The same small model, trained on sequences from 8 to 64 symbols long. In this task every sequence is one short random pattern repeated, so a longer sequence holds more copies of the pattern but nothing new. Grey: what a recent paper's law predicts. *Eight times the length saved three steps out of seventy-five. Extra copies of what an example already contains are not extra data. A task where length adds new content is being tested next.*

**Batch size behaves as the published law says; sequence length does not, here.** The paper's law has two parts: one for batch size and one for context length. Grey bars are the paper's values for large language models; coloured bars are ours, on a small recurrent model at a fixed learning rate. The batch part matches closely. The length part is near zero here, because our task's longer sequences only repeat the same pattern. The law's two halves measure different things: more examples, and more distinct content.

Kill test, fixed before execution: the exponent interval lies entirely outside [-0.93, -0.31]. Anchor, in code: length 16 reproduces T8's batch-64 cell on every receiver. It holds. Prior: low.

Result: the kill test fires

Sequence length	Scored positions per sequence	Step to `0.5`
`8`	`5`	`75.5` `[72.2, 78.9]`
`16`	`13`	`75.3` `[71.8, 78.7]`
`32`	`29`	`73.8` `[70.2, 77.4]`
`64`	`61`	`72.4` `[69.0, 75.8]`
Exponent		`-0.021` `[-0.034, -0.009]`

Twelve times as many scored positions per sequence moves the transition by three steps. The exponent is nearly zero, against the paper's -0.62.

What the design tested, stated plainly

J8's task builds each sequence by repeating one random four-token pattern. A longer sequence therefore adds more positions to score but no new content: every extra position is another copy of the same four tokens. So T9 measured whether more supervised positions of the same pattern speed learning, and they do not. It did not measure whether more context does, in the sense the paper means, where a longer window holds more distinct text. The low prior was written for a different reason (a recurrent model does not look further back when given a longer sequence); the task's repetition is the stronger one and was not noticed until the result was read.

A throwaway check on a task where length does add content -- a random stream copied at a fixed lag of 4, one seed -- moved the step from 231 at length 8 to 104 at length 64, a slope near -0.38. That is one seed and not a result; T11 runs it properly.

What stands

Kill test fires. On J8's task the context exponent is -0.021 [-0.034, -0.009]: more copies of the same pattern per sequence do not bring the transition sooner.
What it says, and does not: repeated positions are not extra data. Examples, not positions, set the pace on this task -- a doubled batch (T8) saves about a fifth of the steps, a doubled sequence about one percent.
Practical reading: padding a training example with more copies of what it already contains buys nothing; count the distinct content, not the tokens.

Limits

The task confounds length with repetition, as above. T11 separates them.
One substrate, one rate, one batch.

FOLLOWED UP 2026-09-27 by T11. On a random stream copied at the same lag, where every extra position is new content, the context exponent is -0.344 [-0.371, -0.316], -0.322 steeper than this record's on the same receivers. The near-zero exponent here was the repetition. Text and numbers unchanged.

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.
confound: A second explanation you did not control for. If bigger models both learn faster and score higher, then 'fast learners score higher' may be entirely about size and not about speed.
exponent: The number in a power law that says how strongly one quantity responds to another. A bigger exponent means a steeper response to the same doubling.
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.
induction: A model's ability to spot that a sequence is repeating and copy what came next last time. A standard early skill in language models.
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.
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.
slope: How steeply one quantity changes as another does. A slope of one between a warning and the event it predicts means the warning shifts exactly in step with the event.