Downloads · 30 days
20
20% of all-time downloads
salih1hf/KobunFormer5-1-8B
KobunFormer5-1-8B is a machine learning model from salih1hf. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
This model is a specialized fine-tune of the Qwen3 8B parameter model, converted to GGUF format using Unsloth. It is expertly trained to translate modern text (across various languages) into Classical Japanese (Kanbun…
Downloads · 30 days
20
20% of all-time downloads
All-time downloads
99
Public
Repo size
13 GB
Likes
0
Public
Click a slice to open those files.
.gguf8.7 GB · 100%
From the Hugging Face model README
This model is a specialized fine-tune of the Qwen3 8B parameter model, converted to GGUF format using Unsloth. It is expertly trained to translate modern text (across various languages) into Classical Japanese (Kanbun-Kundoku style).
KobunFormer5-1-8B employs a unique two-step reasoning approach to ensure high grammatical accuracy and appropriate historical nuances across multiple languages (Tested with Turkish at the moment). It utilizes a thought block to map the linguistic structure before generating the final text:
<think> ... </think> reasoning block.**Example Prompt & Output:**
**System Prompt:** `You are an expert translator specializing in converting texts into Classical Japanese Kanbun-Kundoku style`
**User Input:** `通義千問 çok güzel bir modeldir`
**Model Response:** ```text
<think>通義千問善模型也。</think>
通義千問こそ善き模型なれ。
To get the most accurate structural generation and prevent the model from deviating from its Kanbun-to-Kobun pipeline, the following sampling parameters are highly recommended:
qwen3-8b.Q8_0.gguf (8.71 GB - Q8_0 Quantization)Using llama.cpp via CLI:
llama-cli -m qwen3-8b.Q8_0.gguf --system-prompt "You are an expert translator specializing in converting texts into Classical Japanese Kanbun-Kundoku style" -p "通義千問 çok güzel bir modeldir" --temp 0.7 --top-p 0.8 --top-k 20
(Tested on OSX)
Please note that this was an experimental model.
The model when I last tested was highly unstable (especially on rules like Kakari Musubi, which is hard to get from a Kanbun-> Kobun workflow because there is no standard way of Kakari Musubi in Kanbun Kundoku and its existence depends on the author), but when it answered correctly, it answered really well.