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 the LSTM, Not Settled

By Joshua Bolick, Awareness Software Group ~9 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. On an LSTM, the easy-sum start had looked like a large win, but ordinary training was using settings chosen for a different model. This project has been caught out by that before, so we tuned it.

What we found. Tuning halved the gap. At its best setting, ordinary training solved the hard sum in about 9,200 steps on average against the easy start's 5,400. But its best setting was the most aggressive one we tried, and it was still improving.

Why it matters. So on the LSTM we are not calling it yet. The next test tries faster settings for ordinary training.

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, 48 new training runs of 16000 steps plus 32 re-used from A58. The design and the kill test were committed (bb18a05) before any run.

Program v2 Bucket A, item A59. Decisive computation: analysis/sum_lstm_tuned.py. Output: analysis/sum_lstm_tuned.json.

The question

A58 found on the tiny LSTM that the easy-sum warm-up solves (2, 7) on 15 of 16 seeds at a median 4120 steps while plain training solves 5 -- with the plain LSTM at the GRU's rate (0.003), never tuned for the LSTM. Tuned, does the plain LSTM close the gap?

Tuning ordinary LSTM training halves the gap, and it may close further. Ordinary training on the LSTM at four peak learning rates, against the easy-sum start. Lower bars mean the task was solved sooner; runs that never solved it count as the full 16,000 steps. The best ordinary setting tried (the highest rate) needs about 9,200 steps on average against the easy start's 5,400 -- still slower, but the gap halved and it was still improving, so the result on the LSTM is not settled.

Design: plain LSTM on (2, 7) with a 500-step rate warm-up to peaks 0.002, 0.005, 0.008 (0.003 and the mixed arm re-used from A58), A58's sixteen seeds, 16000 steps; the best plain peak by mean steps to 0.9 against A58's mixed runs. Kill test, fixed before execution: mixed minus the best plain cell includes zero or lies above it; a best peak at a grid edge is reported as not found. Anchor, in code: the 0.003 cell reproduces A58 -- held.

Results

Cell (LSTM, sixteen seeds)	Mean steps to `0.9`	Median	Solved by `16000`
plain, peak `0.002`	`12192`	`15265`	`8`
plain, peak `0.003` (A58)	`12678`	`16000`	`5`
plain, peak `0.005`	`11124`	`16000`	`7`
plain, peak `0.008` (best, grid edge)	`9181`	`6235`	`10`
easier sums mixed first (A58)	`5418`	`4120`	`15`

(Not monotone at the low end: 0.002 does a little better than 0.003. The clear gain is at the higher peaks.)

Mixed minus the best plain cell: -3763.1 [-7359.8, -166.5] steps.

The kill test does not fire -- narrowly, and with the plain run's optimum not found. Tuning the plain LSTM's rate from 0.003 to 0.008 roughly halves A58's gap (-7259 to -3763), and the interval now only just excludes zero. The plain run was still improving at the top of the grid, so a higher rate may close more.

What it says

On the LSTM the speed-up survives tuning so far but is not settled: it is the untuned-baseline trap's signature (a gap that shrinks as the baseline's rate rises), caught before the claim was stated. A61 extends the plain LSTM's rates to 0.012 and 0.016. Until then, the architecture statement is: established on the GRU; leaning, not settled, on the LSTM; absent on the transformer.

What stands

A59: kill test does not fire. Mixed minus the best plain cell (0.008, a grid edge): -3763.1 [-7359.8, -166.5] steps.
The plain LSTM's best rate is not yet found; the LSTM result is not settled.

Limits

Four plain peaks, the best at the edge; sixteen seeds; the mixed arm not re-tuned.

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.
baseline: The thing you compare against. A result without one is not a result.
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.
LSTM: Long Short-Term Memory. An older and larger relative of the GRU, also gated, also for sequences.
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.
settled: A run has settled when it has stopped improving. Measurements anchored to a run's own best score are unreliable until it has, because that best score is still moving.
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.