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

On an LSTM It Holds

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: 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-sum speed-up held on one kind of recurrent network and vanished on a transformer. Was it about recurrent networks in general, or just the one we used? We tried a second recurrent design, an LSTM.

What we found. It held, and more strongly: with the easy start, 15 of 16 runs learned the hard sum, typically in about 4,100 steps; ordinary training managed 5 of 16 in four times as long.

Why it matters. The benefit seems tied to how recurrent networks work: they have to build machinery to hold and combine past symbols, and easy examples help them build it. One caveat we are checking next: ordinary training's settings were not tuned for this model.

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 of 16000 steps on sixteen fresh seeds. The design, a disclosed throwaway and the kill test were committed (a533e6f) before any run.

Program v2 Bucket A, item A58. Decisive computation: analysis/sum_mixture_lstm.py. Output: analysis/sum_mixture_lstm.json.

The question

The modular-sum speed-up holds on the GRU (A44, A53) and does not appear on a one-layer transformer (A55). Is it about recurrence, or about this GRU? The programme's other recurrent architecture is a tiny LSTM.

**The easy-sum start speeds up recurrent networks, not the transformer.** For each kind of model, the average number of training steps the easy-sum start needed to reach the criterion on the hard sum, divided by plain training's (with a learning-rate warm-up). Below 1 means fewer steps. The transformer uses an 80% criterion; the recurrent models 90%. Both recurrent models need roughly half the steps or fewer with the easy start. The transformer needs more, not fewer. The benefit looks tied to how recurrent networks have to hold and combine past symbols.

Design: the LSTM (width 48, one layer); the GRU's task and settings unchanged -- target (2, 7), rate 0.003 -- as a disclosed throwaway showed it inside the LSTM's reach (one seed solved at step 5250); 16000 steps, sixteen fresh seeds. Mixed (2000 steps of (2,3)/(2,4)/(2,5)) against plain with a 500-step learning-rate warm-up; steps to 0.9 (never = 16000). Kill test, fixed before execution: mixed minus plain includes zero or lies above it. Anchor, in code: with the GRU restored, the loop reproduces A40 -- held.

Results

Arm (LSTM, sixteen fresh seeds)	Median steps to `0.9`	Solved by `16000`
easier sums mixed first	`4120`	`15`
plain, learning-rate warm-up	`16000` (budget)	`5`

Mixed minus plain: -7259.4 [-10905.0, -3613.7] steps; ratio of mean steps 0.427.

The kill test does not fire. On the LSTM the mixture solves the task on fifteen of sixteen seeds at a median of 4120 steps, while the plain recipe solves five in the whole budget; the saving is larger than on the GRU and, with eleven plain seeds counted at the budget, understated. (The throwaway's plain seed solving at 5250 was one of the minority.)

What it says

The speed-up is about recurrence. It holds on both recurrent architectures on file (GRU at two widths, LSTM) and not on the transformer. That fits the reading from A49 and A52: a recurrent network must build machinery to hold and combine two past symbols, and the easy sums build it; a transformer can attend to both directly and does not need that help. The reading is still an interpretation; what is measured is the pattern across three architectures. The outside-venue draft's scope is updated to "recurrent networks".

What stands

A58: kill test does not fire. LSTM: mixed minus plain -7259.4 [-10905.0, -3613.7] steps; ratio 0.427; solved 15 against 5.
Architecture pattern: GRU yes (two widths), LSTM yes (larger), one-layer transformer no.

Limits

One LSTM configuration and rate (the GRU's); the plain LSTM was not separately tuned beyond the throwaway, so part of the gap may be the rate suiting the mixture better. 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.

architecture: The blueprint of a model: what components it has and how they connect. Two models can be the same size and completely different architectures.
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.
LSTM: Long Short-Term Memory. An older and larger relative of the GRU, also gated, also for sequences.
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.
transformer: The architecture behind most modern large language models. It uses attention to look at every part of the input at once, rather than reading in order.
width: How many internal numbers a model uses at each layer. The usual way we vary model size in these experiments.