qwen3-1.7B-ko-summary-finetuned-06-12

A fine-tuned Qwen3-1.7B model specialized for abstractive summarization of Korean documents, particularly academic papers. This model was trained on high-quality Korean paper summarization data and enhanced with emotional multi-turn conversation data to expand vocabulary and improve generation quality.

Model Description

• Architecture: Qwen3-1.7B
• Fine-tuning Task: Abstractive summarization
• Training Data: Korean academic paper summaries (e.g., KoreaScience dataset) + Emotional multi-turn conversation data

Key Improvements

1. Resolved Token Repetition Issue: Fixed meaningless token repetition problems from the previous colli98/qwen3-1.7B-ko-summary-finetuned model
2. Structured Summary Format: Improved unstructured summary format issues for better coherence
3. Enhanced Vocabulary: Added emotional multi-turn conversation training data to expand vocabulary range beyond academic papers

Intended Use

• Summarizing long Korean documents—especially research papers—into clear, concise overviews.
• Integrating into research tools, educational platforms, or automated document-processing pipelines.

Performance Evaluation

ROUGE Score Comparison

Metric	Previous Model	Current Model	Improvement
ROUGE-1 Precision	0.357	0.388	+8.7%
ROUGE-1 Recall	0.189	0.174	-7.9%
ROUGE-1 F-measure	0.247	0.241	-2.4%
ROUGE-2 Precision	0.109	0.169	+55.0%
ROUGE-2 Recall	0.058	0.076	+31.1%
ROUGE-2 F-measure	0.075	0.104	+38.7%
ROUGE-L Precision	0.269	0.328	+21.9%
ROUGE-L Recall	0.142	0.147	+3.5%
ROUGE-L F-measure	0.186	0.203	+9.1%
ROUGE-Lsum Precision	0.316	0.319	+0.9%
ROUGE-Lsum Recall	0.168	0.171	+1.8%
ROUGE-Lsum F-measure	0.219	0.223	+1.8%

Performance Analysis

Positive Improvements:

• Overall Precision Enhancement: Improved precision across all metrics, indicating higher quality generated content
• Significant ROUGE-2 Improvement: Major improvement in bigram-level metrics, suggesting more natural and coherent sentence structure generation

Trade-offs:

• Partial Recall Decrease: Slight decrease in recall, particularly in ROUGE-1, suggesting potential missed content from reference texts
• Room for Further Improvement: All metrics remain below 0.4, indicating need for additional performance enhancements

Conclusion: Fine-tuning improved generation quality (precision) while showing slight trade-offs in completeness (recall). The significant ROUGE-2 improvement represents meaningful progress in model performance.

Limitations & Risks

• May produce inaccuracies or hallucinated content.
• Not intended for generating verbatim legal/medical texts or for extractive quotation.
• Users should verify critical facts against original sources.

Installation

pip install transformers safetensors

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("your-username/qwen3-1.7B-ko-summary-finetuned-06-12")
model = AutoModelForSeq2SeqLM.from_pretrained("your-username/qwen3-1.7B-ko-summary-finetuned-06-12")

text = "여기에 긴 한국어 논문 텍스트를 입력하세요..."
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding="longest")
summary_ids = model.generate(
    **inputs,
    max_length=150,
    num_beams=4,
    early_stopping=True
)
summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)
print(summary)

Files in This Repository

.
├── config.json
├── generation_config.json
├── model.safetensors
├── model.safetensors.index.json
├── tokenizer.json
├── tokenizer_config.json
├── special_tokens_map.json
├── vocab.json
├── merges.txt
└── added_tokens.json