Model Card for Model ID
Fine-tune of Qwen2.5-0.5B-Instruct for CoT on Russian.
Model Details
Model Description
- Developed by: sapbot
- Model type: SLM, PoC
- Language(s) (NLP): Russian
- License: MIT
- Finetuned from model: Qwen2.5-0.5B-Instruct
Uses
Direct Use
Check how far does Qwen2.5-0.5B adapt to CoT and Russian language.
Downstream Use [optional]
Generate CoT samples.
Out-of-Scope Use
Anything that needs reliability. This model hallucinates a lot, and dataset has not been filtered for anything specified. There is straight up garbage in dataset, and no specific censoring. Behavior is awaited to be same as in samples of dataset.
Bias, Risks, and Limitations
Main risk is unreliable information. This model can straight up imagine new facts. It's always picking up some strange things from context. Like, when I fed it date in Unsloth Studio, it started to think that I give it questions about the date. Also, it isn't trained on languages other then Russian.
Recommendations
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations. And don't use it for anything in production. It's basically too raw. Either finetune your own model on the same dataset or straight up pick something more reliable (I would recommend the Qwen3.5, they have pretty solid CoT while being almost the same size).
Training Details
Training Data
Trained on synthetical CoT on Russian, generated by Gemma 3 4B. Questions and answers are from YandexQ, the UGC questions platform. Not all answers are verified.
Training hardware and time
Trained for 9m 25s on RX 6600 (8GB VRAM 16GB RAM)
sapbot@sapbot-rx6600:~$ fastfetch
`.:/ossyyyysso/:. sapbot@sapbot-rx6600
.:oyyyyyyyyyyyyyyyyyyo:` --------------------
-oyyyyyyyodMMyyyyyyyysyyyyo- OS: Kubuntu 26.04.1 LTS (Resolute R4
-syyyyyyyyyydMMyoyyyydmMMyyyyys- Kernel: Linux 7.0.0-30-generic
oyyysdMysyyyydMMMMMMMMMMMMMyyyyyyyo Uptime: 6 hours, 40 mins
`oyyyydMMMMysyysoooooodMMMMyyyyyyyyyo` Packages: 3018 (dpkg), 10 (snap)
oyyyyyydMMMMyyyyyyyyyyyysdMMysssssyyyo Shell: bash 5.3.9
-yyyyyyyydMysyyyyyyyyyyyyyysdMMMMMysyyy- Display (2276WM): 1920x1080 in 22",*
oyyyysoodMyyyyyyyyyyyyyyyyyyydMMMMysyyyo Display (AAA): 1366x768 in 26", 60 ]
yyysdMMMMMyyyyyyyyyyyyyyyyyyysosyyyyyyyy DE: KDE Plasma 6.6.6
yyysdMMMMMyyyyyyyyyyyyyyyyyyyyyyyyyyyyyy WM: KWin (Wayland)
oyyyyysosdyyyyyyyyyyyyyyyyyyydMMMMysyyyo WM Theme: Breeze
-yyyyyyyydMysyyyyyyyyyyyyyysdMMMMMysyyy- Theme: Breeze (Light) [Qt], Breeze ]
oyyyyyydMMMysyyyyyyyyyyysdMMyoyyyoyyyo Icons: breeze [Qt], oxygen [GTK2/3/]
`oyyyydMMMysyyyoooooodMMMMyoyyyyyyyyo Font: Noto Sans (10pt) [Qt], Noto S]
oyyysyyoyyyysdMMMMMMMMMMMyyyyyyyyo Cursor: breeze (30px)
-syyyyyyyyydMMMysyyydMMMysyyyys- Terminal: konsole 25.12.3
-oyyyyyyydMMyyyyyyysosyyyyo- CPU: Intel(R) Xeon(R) E3-1270 V2 (8z
./oyyyyyyyyyyyyyyyyyyo/. GPU: AMD Radeon RX 6600 [Discrete]
`.:/oosyyyysso/:.` Memory: 6.76 GiB / 15.57 GiB (43%)
Swap: 506.44 MiB / 512.00 MiB (99%)
Disk (/): 53.82 GiB / 218.71 GiB (24
Disk (/media/sapbot/hdd): 112.29 Gi4
Disk (/media/sapbot/steam): 1.59 Tis
Local IP (enp5s0): 192.168.1.2/24
Locale: ru_RU.UTF-8
Environmental Impact
- Hardware Type: RX 6600 8GB
- Hours used: 0.15
- Cloud Provider: Private Infrastructure
- Compute Region: Russia
- Carbon Emitted: 16.9 gCO₂
Technical specifications
Software
Latest at the time Unsloth 2026.11.4
Framework versions
- PEFT 0.18.1