qwen38-et-adapters

qwen38-et adapters — Estonian post-training for Qwen3.8-27B

Three LoRA adapters from a fully documented Estonian post-training project (single RTX 5090). Method, scripts, eval and lessons: github.com/pertlomp/qwen38-et

⚠️ Read this before quoting the scores. The 200-task Estonian set was consulted after every training round and used to choose the next training batch. That makes it a development set, not a held-out test, and the scores below are optimistically biased by an unknown amount. They are not comparable to numbers other models report on their own benchmarks. No independent blind evaluation has been done. The most defensible result here is the perplexity drop, because nothing was tuned against it. Full write-up of this and of three measurement errors found later: PARANDUSED.md.

*200-task Estonian set, 151 auto-scored (47 need human judgement, 2 are code tasks the scorer skips). GPT scored 81.7 and Gemini 84.3 on the same set, but they saw it cold while this model was tuned against it over 13 rounds, so the comparison is biased in this model's favour and the set itself was built from the base model's own errors.

CPT details: 110.2M tokens actually reached the model (3364 steps × 16 accumulation × 2048 context) out of a ~141M-token corpus; the difference was lost to truncation. Replay was 3.1% of words. 35.3 h at ~870 tok/s, 450 W.

Usage: apply over Qwen3.8-27B with PEFT, or merge. Important: trained no-think — always use think=false / enable_thinking=False. For GGUF/Ollama use Q5+ quantization (Q4 measurably degraded output in our test, though we did not isolate whether quantization itself was the cause) and set RENDERER/PARSER explicitly.

Eesti keeles: kolm LoRA adapterit eesti keele järeltreeningu projektist. Metoodika, skriptid ja õppetunnid GitHubis. Treenitud think-režiimita — kasuta alati think=false. NB: 86,1% on arendusmõõt, mitte sõltumatu testitulemus; vt PARANDUSED.md.

Non-commercial work. Contact: [email protected]