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

Memorised, Never Generalised

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 decides when a model learns, and can you change it? – Settings dominate, data barely matters, and there is a brief window before the jump in which interrupting the model is unusually costly. Timing can be delayed but not brought forward.

In plain English

What we asked. 'Grokking' is a famous effect: train a network over and over on a small fixed set of arithmetic problems, and long after it has memorised them it suddenly learns the real rule. All our experiments so far used fresh problems every step, so we tried a small fixed set.

What we found. Our model memorised the problems within about a hundred steps and never solved a new one in ten thousand, whatever the setting. A standard trick that helps grokking, weight decay, slowly wore away the memorised answers without anything better taking their place.

Why it matters. The sets were probably far too small: grokking happens near the amount of data a model needs to find the rule. The next experiment moves up to that point.

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, 24 training runs of 10,000 steps. The design and kill test were committed (0615361) before any run; one streaming throwaway run was made first and is disclosed.

Program v2 Bucket E, item E5. Decisive computation: analysis/grokking_bridge.py. Output: analysis/grokking_bridge.json. Reproduce with python analysis/grokking_bridge.py (about an hour on a throttled laptop CPU); --reuse re-derives every endpoint from the saved curves.

The question

Every run in this programme drew fresh data at every step, so memorising the training set was impossible by construction. Grokking (Power et al., 2022, arXiv 2201.02177) is the best-studied sudden-learning phenomenon in the literature: on a small fixed training set of a modular-arithmetic task, with weight decay, a network first fits its training set and only much later generalises. E5 asks whether this rig produces it.

**Small training sets are memorised at once and never generalised.** A small model trained over and over on a fixed handful of addition problems, four handful sizes. Solid lines: how many of the problems it was trained on it gets right. Dashed: how many new problems it gets right. It learns the problems it was shown within about a hundred steps and never solves a new one; the dashed lines stay at guessing level for 10,000 steps. The famous 'grokking' delay did not appear with training sets this small.

Task: modular sum (x[t-2] + x[t-4] mod 32), J8's width-48 GRU, AdamW at 0.005; a streaming throwaway run learned it by step 2000. Grid: fixed pools of 8, 16, 32, 64 sequences (12 scored sums each), weight decay 0 and 1.0, three seeds, 10,000 full-batch steps. A run groks if training accuracy reaches 0.99 while held-out accuracy is below 0.5, and held-out accuracy later reaches 0.9 at least three times as late.

Kill test, fixed before execution: no cell has two of three seeds that grok.

Result: the kill test fires

Pool	Weight decay	Grokked	Memorised by step	Best held-out accuracy (seed `8101`)
`8`	`0` / `1`	`0/3` / `0/3`	`50`	`0.034` / `0.035`
`16`	`0` / `1`	`0/3` / `0/3`	`50`	`0.033` / `0.036`
`32`	`0` / `1`	`0/3` / `0/3`	`50`-`100`	`0.036` / `0.041`
`64`	`0` / `1`	`0/3` / `0/3`	`100`	`0.033` / `0.038`

(Chance is 1/32 = 0.031.)

Every run memorises its training set within 50-100 steps and none ever generalises: held-out accuracy stays within a point of chance for all 10,000 steps. With weight decay the memorised solution also erodes: in every one of the twelve weight-decay runs, training accuracy later falls well below 0.99 (lowest values after step 1000 between 0.26 and 0.65), and nothing general replaces it.

Why this is not the end of the question

The bounded search the kill test named contains no grokking. Two features of the design make that a statement about this search rather than about the rig, and both are stated here rather than discovered later:

The pools are far below what the task needs. The streaming run that learned the task saw about 64,000 sequences by step 1000; the largest pool here holds 64. Grokking in the literature sits at the edge of data sufficiency, where a network can eventually find the general rule from the data it has; these pools may be too far below that edge for any delay to end.
A recurrent model can memorise whole sequences. With 8-64 distinct sequences it can key on each sequence's opening tokens and never learn the sum at all.

E10 tests the edge directly: larger pools, up to a size where the model generalises, and whether a delay appears just below it.

What stands

Kill test fires. Pools of 8-64 sequences, with or without weight decay: memorised in 50-100 steps, never generalised in 10,000.
Weight decay erodes memorisation here without producing generalisation.

Limits

One task, one width, one rate, 10,000 steps. Power et al.'s delays can run to 10^5 steps and more.
The pool range sits far below the streaming run's data.

FOLLOWED UP 2026-09-27 by E10. Larger pools (256-16384) find the edge: with weight decay the model generalises from 1024 sequences, without it from 4096, and in every such run it generalises no later than it memorises. No grokking in either record. Text and numbers above 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.
AdamW: A widely used training algorithm that adapts how big a step it takes for each individual parameter. The default in most modern AI training.
grokking: When a network that has memorised its training examples suddenly, much later, learns the general rule and starts getting new examples right.
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.
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.
weight decay: A common training setting that gently pulls a model's internal numbers toward zero, used to stop it over-fitting.
width: How many internal numbers a model uses at each layer. The usual way we vary model size in these experiments.