Skip to content
Projects

CLOVA Note Title Generation Model

NAVER Cloud Document Intelligence internship project. Built summary-based title generation for CLOVA Note and improved weak label quality via positive filtering SFT.

2022.12 — 2023.02 NAVER / NAVER Cloud AI Research Engineer (Intern)
NLP HyperCLOVA LoRA Weak Labeling

Overview

During an internship on the NAVER Cloud Document Intelligence team, I developed an automatic title generation model based on CLOVA Note meeting summaries. Applied LoRA tuning on HyperCLOVA/LaRva backbones and systematically improved weak label quality through multi-criteria positive filtering.

Key Achievements

  • Designed a multi-criteria filtering pipeline with 4 criteria:
    • ROUGE-1 precision >= 0.75
    • Sentence-encoder-based semantic similarity >= 0.9
    • 2 additional quality filters applied
  • Reduced bullet-format error rate from 6.75% to 3.50% (-48%) in the first SFT round
  • Further reduced bullet-format errors from 12 to 1 (-92%) in subsequent experiments
  • Improved train filtered count from 83.71% to 95.98%
  • Performed OOD validation (news articles, course syllabi, multilingual data)
  • Launched as the official CLOVA Note “AI Summary Subtitle Support” feature 3 months after the internship ended

Design Constraints & Decisions

  • Had to improve quality from weak labeling alone without large-scale manual annotation
  • This was late 2022 — before DPO (Direct Preference Optimization) existed — so improving model preferences had to be achieved purely through SFT + automatic data curation
  • Chose an iteration structure retaining only positive data for retraining, rather than stopping at rule-based quality constraints
  • Separated tone & manner from length constraints, iteratively tuning metric combinations in follow-up experiments

Technical Approach

  • Used HyperCLOVA/LaRva large language models as backbones with efficient fine-tuning via LoRA/PEFT
  • Applied positive filtering SFT methodology to improve weak label data quality: selected only samples passing quality criteria from first-round SFT model outputs for second-round SFT
  • Multi-axis filtering combining ROUGE-based lexical overlap and sentence-encoder-based semantic similarity for noise removal
  • Validated generalization performance on OOD data including news articles, course syllabi, and multilingual inputs
  • Contributed a PR to the internal SFT finetuning platform, enabling comparison of multiple metrics at once during intermediate validation

Tech Stack

HyperCLOVA, LaRva, LoRA/PEFT, ROUGE, Sentence-Encoder, Streamlit, Python