Downloads · 30 days
6
5% of all-time downloads
qwen3.5-2b-wmt26-sorbian
mkhasykov/qwen3.5-2b-wmt26-sorbian
qwen3.5-2b-wmt26-sorbian is a machine learning model from mkhasykov. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Primary submission model of team HeyBusan (OCELoT 123-127, constrained/open weights). Qwen3.5-2B, full-parameter fine-tune for the five shared-task tasks (MT de/hsb/dsb, QA, spell checking, grammar checking, maths rea…
All-time downloads
127
Public
Parameters
1.9B
3.8 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors3.8 GB · 99%
At a glance
- License
- apache-2.0
- Model type
- qwen3_5_text
- Access
- Public
- Created
- Jul 20, 2026
- Updated
- Jul 21, 2026
- SHA
- 4407a38d
Base models
Model card
From the Hugging Face model README
HeyBusan — WMT26 Multitask LLMs with Limited Resources, Sorbian track
Primary submission model of team HeyBusan (OCELoT #123-127, constrained/open weights). Qwen3.5-2B, full-parameter fine-tune for the five shared-task tasks (MT de/hsb/dsb, QA, spell checking, grammar checking, maths reasoning), merged 0.9/0.1 in weight space with the base model (the base admixture restores maths/QA ability lost to multitask SFT).
Live test scores: MT chrF 72.3 / QA 49.4 / SC 69.7 / GC 76.2 / MR 28.4.
Training data: official task data (MR dev excluded — PolyMath compliance), decontaminated WMT20-22 parallel corpora, FineWeb-2 hsb/dsb monolingual, public instruct mix, soblex-based synthetic SC/GC, STaR-verified German GSM8K-train chains. Decoding at submission: MT beam=4, SC/GC greedy, QA loglik-argmax, MR combined direct+translated self-consistency (16 samples/view @T0.7).
This repo always holds the team's current best model; earlier versions live in the git history. Contact: [email protected]
- Type
- qwen3_5_text
- License
- apache-2.0
- Languages
- hsb, dsb, de
- Created
- Jul 20, 2026
- Updated
- Jul 21, 2026