One GPU, one person, every step measured. I taught the open language model Qwen3.8-27B proper Estonian on a single RTX 5090. Here are the results with their denominators, the mistakes with their price, the model and code in the open, and a textbook you can download right now.
Open language models speak English well and Estonian carelessly. They inflect wrongly, mix up verb government and write sentences no Estonian would write. Paid models are better, but behind them are someone else's decisions and someone else's price. I wanted to know how far one person gets with one ordinary gaming PC, if every step is measured.
I chose Qwen3.8-27B as the base, an open-weights model with 27 billion parameters, and trained it with QLoRA on a single RTX 5090 with 32 GB of memory. The project started on 22 August 2026. In nine days I ran 34 saved training runs: first 13 targeted skill rounds, where each round's material came from the model's own mistakes in the previous round, then 110 million tokens of continued pretraining on edited Estonian prose, and finally surgical repair rounds on top of that.
The most expensive lesson was not technical. It was about measurement. My 200-task Estonian test was meant as a locked check, but I looked at it after every round and chose the next batch by it. By definition that makes it a development set, not a test. The number that comes out of it is biased upwards, and I do not know by how much. That is exactly why every number on this page carries its denominator and its limitation next to it, in the same type size.
Second lesson: a small targeted dose beats a large general one. 720 examples aimed at one specific error gave more than 81,268 general examples. And third: the model's own mistakes are the best training material there is. The same data as one big batch scored 9 points lower than 13 rounds where each round looked at the previous one's errors.
Estonian is a small language. Its training must be publicly available and the work must carry forward into future models, not stay on one person's hard drive. This work has no economic weight for me, neither direct nor indirect. The code is MIT, the adapters Apache 2.0, the reports CC BY 4.0. The mistakes are public too, because they teach more than a clean final result.
A number without a denominator cannot be checked. Here are the four measurements I made, and next to each one what it says and what it does not say. The most defensible is perplexity, because no round was tuned against it.
All numbers, checkpoints and the three measurement errors found later are in the repo file PARANDUSED.md. Read it before quoting any number.
Data inventory, a 200-task Estonian test, base model measured: 61.7 %.
Rule-generated inflection examples, round-tripped through Vabamorf. From there, round after round from the model's own errors.
84.7 % on the dev set, but EstQA fell below base. Three different models review the plan and find the gaps.
110 M tokens of prose, perplexity −31 %. Surgical rounds on top, EstQA recovers to 93.6. LIHV2 is the final state.
Adapters, GGUF and corpus on Hugging Face. Cross-checking finds three measurement errors; every public artefact is corrected.
Four errors at once. The pure language layer alone made the model 5.6 points worse on the dev set. Rebuilding, behind a rule gate.
Every mistake here is measured, not generalised. Without a price a lesson is not convincing and cannot be handed on. None of them raised an error. Every time, the script said "done".
| No | Mistake | Measured price | How it was found |
|---|---|---|---|
| 01 | Packing assumed 4 characters per token. In Estonian it is 2.8. | About 22 % of the corpus was silently truncated and never reached the model. | A second model asked in cross-review whether chunk counts and token counts matched. They did not. |
| 02 | Replay data was declared in the code but never used. | The pure language layer took the model from 61.7 to 56.1, a 5.6-point forgetting. | By checking an outside recommendation against the code, not against the documents. |
| 03 | Perplexity was measured on one genre only. | Modern news text got 16.4 % worse and nobody saw it. | A multi-genre baseline before the next training run. |
| 04 | The "locked test" was actually a development set. | 86.1 % is not a test result. The bias is of unknown size. | Cross-model review: three independent models read the textbook. |
| 05 | Data was uncleaned: wiki markup, OCR noise, template text. | About 24 % of the corpus was not whole Estonian text. | A seven-criteria data audit on 4 September. |
| 06 | The learning rate decayed to zero inside every block instead of across the run. | About 20 % of every block trained on nothing. | By reading the trainer state file, not the log. |
| 07 | A chat variant was trained on raw text without instruction examples. | Loss of instruction skill, part of mistake 02. | By reading the EstLLM paper, the only published Estonian CPT mix. |
On 4 September I withdrew the language layer. Not because the result was bad, but because I could no longer say which part of the result was the language layer and which part was chance. If one error had been enough, four at once leave no choice.
Now everything we know about training an Estonian model lives in one repository and one rules file. A rule the machine does not check is a suggestion. Suggestions get forgotten.
36 rules, 33 of them mandatory, each born from a measured mistake and each with its price. The rules file is not documentation. A pre-check reads it, and training does not start if any mandatory rule is unmet.
Packing uses the tokenizer, not character counts. Replay data is really in the mix. Perplexity is measured on at least five genres, in every block. And the validation set exists before training, not after.
The champion checkpoint in one file, Ollama Modelfile included. Trained without think mode: use think=false. Q5 or better; Q4 produced measurably worse text.
Open model card ↗The final champion (language layer + surgical rounds), the pure language layer as a research object, and the 13-round skills path without a language layer. Each with its own measurements and warnings.
Open adapters ↗11,011 entries, each with its source and licence. Entries with an unresolved licence are withheld until cleared. Round-tripped through the Vabamorf morphological analyser.
Open dataset ↗50 numbered scripts and 26 orchestration chains: data processing, rule-based generators, QLoRA training, locked eval and external benchmarks. Reports, council synthesis, the corrections file.
Open repo ↗Three measurement errors, the development-set warning and the licence position, written down before anyone else had to ask. The numbers look weaker after the correction. That is correct.
Read the corrections ↗What was measured on which checkpoint, by category. A phrase probe on 378 unseen cells, cloud models as context on the same EstQA sample, and what must not be read into it.
Open report ↗This book grew out of this very project: collecting data, running experiments, making mistakes and fixing them. 20 chapters, exercises with answers, a glossary and two runnable labs. You do not need to know a single acronym at the start. Written in Estonian.
The recipe is sequential because every step depends on the previous one's measurement. Script numbers refer to files in the repo.
Aimed at your language's weaknesses. And keep it locked: if you look at it after every round, it is no longer a test.
Where does the model fail most? That is where the first dose goes, not where the data is most plentiful.
Plus a morphological analyser as the check. For Estonian, Vabamorf. The shape of the example must match the shape of the use.
2 to 3 hours each. Read the errors before collecting more text. 130 examples for one error gave +2.1 points.
On-policy. 9,543 foreign preference pairs made the model 6 points worse; 244 of its own pairs added a point.
CPT on quality prose gives the language, SFT gives the skills. Replay in the mix, validation set before, not after.
EstQA, MMLU-et, HumanEval. And check the packaged model separately: quantisation and think mode break things silently.
A rule the machine does not check is a suggestion. Turn every mistake into a rule that the pre-check actually reads before the next run.
I do not talk about AI from the sidelines. Every day I run several AI systems that do real work: they write and check code, keep my knowledge base current and run Evoluna's matching. The same discipline as in the language model project applies to each of them: one model builds, another reviews, and the measured result beats the feeling.
This is also the backbone of my mentoring work. When I talk with a leader about AI-era decisions, it does not come from a book. It comes from systems that are running right now. See the full AI practice on the home page →
Data with a clear licence. An independent blind test that would make the numbers truly defensible. Or simply a question about how to start with your own language or domain. I answer personally.
Write: info@pertlomp.com