pertai/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

Adapter	What it is	Score*
KROON-tsempion-86-1/	Final champion: 131M-token CPT on edited Estonian prose + surgical skill rounds. Fiction perplexity −31% vs base, HumanEval 85.4%, EstQA reading F1 93.6 (above Claude Fable's 92.4 on the same sample).	86.1%
cpt1-puhas-keelekiht/	Pure CPT language layer (rank 32), no skill training — a research object: what does 131M tokens of edited prose alone do?	ppl 22.2→15.4
ring12-oskuste-tsempion/	13 iterative targeted SFT rounds + on-policy DPO (rank 16), no CPT.	85.3%

*Locked 200-task Estonian eval (in the GitHub repo); GPT scored 81.7, Gemini 84.3 on the same test.

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 damages CPT-shifted weights) 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.

Non-commercial work. Contact: pertlomp@gmail.com