Verified, then logged.
Sun 22 Nov 2026 · 00:46 KST / 15:46 UTC한국어 | ENSearch /

Breaking01 / AI

Linear merging lifts Turkish model's maths score to 0.63

Researchers reported that data quality mattered more than training volume in fine-tuning.

A researcher seen from behind compares mathematical solution paths and merged model-weight heatmaps on two monitors in a university computer lab. AI-generated image

Researchers reported a GSM8K maths benchmark score of up to 0.63 for a Turkish-language model using linear merging in a paper published on 10 Nov. The figure was reported by the researchers.

According to the paper, the study used Turkish-Llama 8B-DPO. It combined LoRA fine-tuning, which adjusts a subset of model weights, with model-weight merging and inference that explores alternative reasoning paths.

The researchers used Turkish translations of MetaMath, NuminaMath and TigerLab for training. Mix-30k performed best among the fine-tuned models, and the paper reported that data quality mattered more than dataset size.

The paper also reported improved reasoning quality when the model explored multiple reasoning paths. Average reasoning-quality scores assessed by a language model rose in the researchers' reported results.

Among the merging methods, linear merging produced the largest gains, according to the paper. Ties delivered limited gains, while Breadcrumbs and Slerp reduced performance. The researchers identified machine translation, the limited scale of the experiments and randomness in the search process as limitations.

The abstract does not disclose scores for individual search methods or variation across repeated runs.