---
language:
- ko
- zh
language_bcp47:
- zh-Hant
license: other
library_name: peft
pipeline_tag: text-generation
tags:
- joseon
- hanmun
- hanja
- korean-translation
- lora
- qwen
- historical-translation
base_model: Qwen/Qwen3.6-27B
datasets:
- custom
metrics:
- accuracy
---

# Joseon Level 4 Qwen3.6 27B LoRA

This is the current best Level 4 translation adapter for the Joseon-to-Day
project. It is a PEFT LoRA adapter trained on top of `Qwen/Qwen3.6-27B`.

The model translates annotated Joseon historical Hanja/Hanmun source into modern
Korean prose. Inputs are expected to include compact structured hints from the
project pipeline, such as Level 1 spans, Level 2 훈음 candidates, and office/title
term hints.

## Model Type

- Base model: `Qwen/Qwen3.6-27B`
- Method: bf16 LoRA
- LoRA rank: 32
- LoRA alpha: 64
- LoRA dropout: 0.05
- Target modules: `q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`,
  `up_proj`, `down_proj`
- Sequence length used for this run: 2048

This repository contains the LoRA adapter only, not the full 27B base model.

## Evaluation Snapshot

Partial dev audit from the local project evaluation:

| Metric | Value |
|---|---:|
| person reading accuracy | 1.0000 |
| office reading accuracy | 1.0000 |
| entity F1 | 1.0000 |
| institutional term accuracy | 0.9709 |
| action accuracy | 0.8800 |
| polarity accuracy | 0.8800 |
| Hanja output rate | 0.0000 |
| Korean-only output rate | 1.0000 |
| historical integrity composite | 0.8642 |

Included diagnostics:

- `dev_sample8_eval.json`
- `dev_partial_eval.json`
- `launch.log`
- `readiness.json`

## Intended Use

Use this adapter inside the Joseon-to-Day MVP translator pipeline. The model is
intended for structured historical translation where the source is accompanied
by reading/entity/institution hints.

## Limitations

This is not a generic Traditional Chinese to Korean translation model. It was
trained for a project-specific annotated prompt format and should be evaluated
carefully before use in scholarly or production workflows. Event/action fidelity
is improving but remains a known evaluation target.

## Loading

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3.6-27B"
adapter_id = "suhjae/joseon-level4-qwen36-27b-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)
```