🏆 CPDC 2023: Winners Announcement

Hello Participants,

We’re excited to announce the tentative winners of the Commonsense Persona-Grounded Dialogue Challenge (CPDC) 2023! A big thank you to all who participated, experimented, and pushed the boundaries of persona-grounded dialogue systems. The challenge received hundreds of submissions, and after a rigorous evaluation, here are the top performers in each task.

🧠 Task 1: Commonsense Dialogue Response Generation
🥇 1st Place: @ni_kai_hua – $15,000
🥈 2nd Place: @wangzhiyu918 – $7,000
🥉 3rd Place: Team justsnail (@jiayu_liu, @kevin_yan) – $3,000

📚 Task 2: Commonsense Persona Knowledge Linking
🥇 1st Place: @biu_biu – $5,000
🥈 2nd Place: Team test_team (@wangxiao, @yiyang_zheng) – $3,000
🥉 3rd Place: @TieMoJi – $2,000

🔦 Winner Spotlight

🌟 @ni_kai_hua – 1st in Task 1

Background: AI graduate from the University of Leeds with experience at Augmentum and CareerBuilder.
Expertise: AI, deep learning, language dynamics.

🏗️ Solution Highlights

• Fine-tuned an LLM on curated persona datasets
• Crafted precise prompts embedding persona traits
• Tuned attention windows for dialogue coherence
• Built a custom evaluation loop for feedback refinement
• Integrated ethical safeguards to ensure responsible output

🔧 Key Techniques
Fine-tuning + prompt engineering synergy, privacy-aware generation, fallback mechanisms for uncertainty.

👉 Read More

🌟 @wangzhiyu918 – 2nd in Task 1

Team: Zhiyu Wang, Puhong Duan, Zhuojun Xie, Wang Liu, Bin Sun, Xudong Kang, Shutao Li
Background: PhD candidate at Hunan University focused on vision-language and multi-modal LLMs.

🏗️ Solution Highlights

• Inspired by ChatGPT-style prompting techniques
• Applied clarity, conciseness, and relevance principles
• Designed and tested high-performing prompt templates

🔧 Key Techniques
Prompt-focused strategy refined with GPT-4, iterative design loop.

👉 Read More

🌟 Team justsnail (@jiayu_liu, @kevin_yan) – 3rd in Task 1

Background: Kaidi Yan is a software engineer focused on C++ and LLMs.

🏗️ Solution Highlights

• Minimalist, strategic prompts to simulate dialogue flow
• Special prompt design for openers
• Single unified user prompt for full dialogue history
• Lightweight post-processing for fluency and completeness

🔧 Key Techniques
Strategic brevity + adaptive prompt design for generalisation.

👉 Read More

🌟 @biu_biu – 1st in Task 2

Background: NLP practitioner with a focus on dialogue systems and commonsense reasoning.

🏗️ Solution Highlights

• Ran baseline evaluation on hidden test sets
• Augmented dataset using GPT-3.5-Turbo to generate 20,000+ synthetic conversations
• Fine-tuned DeBERTa-V3 with thorough hyperparameter tuning
• Modelled head/tail fact prediction paths separately and jointly

🔧 Key Techniques
Synthetic data generation + modern transformer architectures for knowledge grounding.

👉 Read More

🌟 @TieMoJi – 3rd in Task 2

Background: AI algorithm engineer at NetEase Games (NLP and deep learning).

🏗️ Solution Highlights

• Translated structured persona facts into natural language triples
• Framed task as sentence-triple correlation
• Used multi-prompt training and model fusion for robust learning

🔧 Key Techniques
Template-based structuring, mixed precision training, and class-balanced resampling.

👉 Read More

🌟 @yiyang_zheng – 2nd in Task 2

Team: Yiyang Zheng (Shanghai University), Yingwu Yang (ML in finance)

🏗️ Solution Highlights

• Fine-tuned Phi-2 for efficient multi-turn reasoning
• Focused on implicit persona alignment and ambiguous fact handling
• Balanced between fluency and commonsense-grounded accuracy

🔧 Key Techniques
Compact LLM selection (Phi-2), subtle context reasoning, and fine-tuned generalisability.

👉 Read More