arXiv cs.CL· Tohir Saidzoda, Madina Abdusalomzoda, Nasimjon Rajabov, Akhrorkhon Ikromov, Komron Giyosiev, Abubakr Abdullozoda, Amin Jalilov, Jonibek Kishvarov, Rahmonali Sharipov, Mariyamkhon Shodyeva·· 11 小时前AI 评分34
将 Qwen3 适配到塔吉克语（0.6B 至 8B）：文本损失下降不等于理解能力提升
Adapting Qwen3 to Tajik from 0.6B to 8B: Lower Text Loss Did Not Mean Better Understanding
AI 导读
在塔吉克语新闻和书籍上训练的 LoRA 适配器将 Qwen3（0.6B 至 8B）的塔吉克语文本损失降低了 0.60 至 1.45 nats/token，但在可测量的阅读理解上均无提升：Belebele 上 1.7B 和 8B 适配器准确率不变，4B 适配器反降 3.7 分。
正文 · 原文
Abstract:Small LoRA adapters trained on Tajik news and books lowered Qwen3's Tajik text loss at every size we tried, from 0.6B to 8B, by 0.60 to 1.45 nats per token, but none of them improved Tajik reading comprehension where it can be measured. On Belebele the 1.7B and 8B adapters left accuracy unchanged and the 4B adapter lowered it by 3.7 points; at 0.6B the score rose by 4.2 points but sits near chance and coincides with a change in which answer letter the model prefers, so we do not read it as understanding. At 8B, continued pretraining on about 375 million tokens of raw Tajik text was followed by a Belebele gain of 9.9 points. After a further supervised fine-tuning stage, our main model, e3, is 11.4 points above Qwen3-8B on Belebele; its English MMLU changed by -1.2 points, and a loss larger than 3.3 points is excluded. The same heavy recipe applied to Qwen3-0.6B-Base brought its text loss close to that of the 8B model after books-only pretraining, while its multiple-choice answers stayed near chance and its English MMLU fell by 7.6 points. We report 15 trained models and their 5 starting models from June to October 2025, scored in 2026 on one harness with paired intervals, and we report what went wrong: lost training records, a thinking-mode dataset with almost no reasoning in it, later branches whose training data was never documented, and a 4-bit release that contains no transformer layers.
| Comments: | 8 pages, 3 tables, 1 figure. Technical report |
| Subjects: | Computation and Language (cs.CL) |
| Cite as: | arXiv:2611.06062 [cs.CL] |
| (or arXiv:2611.06062v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2611.06062 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Tohir Saidzoda Mr [view email]
[v1]
Wed, 4 Nov 2026 07:08:21 UTC (46 KB)
来源：arXiv cs.CL · arxiv.org