Model Card for gj5520/kkachi60_en2ko

Model Description

kkachi60_en2ko is a sequence-to-sequence Transformer model fine-tuned for English→Korean translation. It is based on the original checkpoint at step 300k (checkpoint-50003) and further adapted with a combination of dataset A and BplusD for improved fluency and adequacy.

• Architecture: AutoModelForSeq2SeqLM (likely based on T5 or mBART architecture)
• Checkpoint: checkpoint-50003 from chkpt_prime_300k_A_BplusD
• Tokenizer: AutoTokenizer matching the model architecture

Intended Use

• Primary use case: Translating English text to Korean for research, prototyping, and evaluation tasks.
• Language pair: English (source) → Korean (target)
• License: Please refer to the Hugging Face hub license setting for usage rights.

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

model_name = "gj5520/kkachi60_en2ko"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

def translate(text: str):
    prefix = "translate English to Korean: "
    inputs = tokenizer(prefix + text, return_tensors="pt", truncation=True)
    outputs = model.generate(
        **inputs,
        max_new_tokens=100,
        num_beams=5,
        no_repeat_ngram_size=3,
        early_stopping=True,
    )
    return tokenizer.decode(outputs[0], skip_special_tokens=True)

sample = "This is a test sentence for translation."
print(translate(sample))

Evaluation Results

Performance was measured on FLORES-200 and WMT24 benchmarks using COMET, BLEU, ChrF, and spBLEU metrics.

Benchmark	COMET (mean)	BLEU (corpus)	ChrF (corpus)	spBLEU (corpus)
FLORES-200	/	/	/	/
WMT24	/	/	/	/

(Insert actual scores here after running benchmark_len.py.)

Training Data

• FLORES-200 (test split)
• WMT24++ English–Korean (en‑ko_KR config)

Training Procedure

• Fine-tuned for 300k steps on mixed data A+BplusD.
• Beam search with num_beams=5, no_repeat_ngram_size=3.
• Generation prefix: "translate English to Korean: ".

Limitations and Biases

• May underperform on very long sentences or domain-specific jargon.
• Potential biases inherited from training corpora.
• Evaluate outputs critically, especially for sensitive content.

Citation

If you use this model in your work, please cite:

@inproceedings{kkachi60_en2ko,
  title={{kkachi60\_en2ko}: English-to-Korean translation model},
  author={Gj5520},
  year={2025}
}

---

This model card was automatically generated based on the model training and evaluation pipeline.