Building a Text2cypher Model
summary
Merged and cleaned around 25 public and internal Cypher datasets into a unified English corpus, producing ~40k training / ~5k test pairs with question, schema and Cypher fields; benchmarks run on open and closed foundation models (Hugging Face, GPT, Gemini) plus a Neo4j fine‑tuned model, evaluated with BLEU and exact execution metrics and embedding similarities, finding GPT-family and Neo4j Thomas fine‑tuned model lead overall; execution tests reveal syntax failures and data‑leakage/overfitting risks; demo shows foundation models sometimes return syntactically incorrect Cypher and fine‑tuning plus curated data improves results.
AI-written summary · about
chapters (11)
- 0:00Introduction to Text2Cypher
- 1:13What is Text2Cypher
- 3:01Methodologies for building models
- 3:22Data collection and preparation
- 5:26Training and test dataset
- 6:08Model evaluation process
- 7:44Evaluation results and findings
- 9:14Benchmarking additional models
- 10:38Challenges and future work
- 11:47Live demo of Text2Cypher
- 13:37Conclusion and Q&A
related talks
23mImproving Text2Cypher: Lessons from Schema Filtering and Hard Example Selection
14mText2Cypher Beyond English: Initial Evaluation with Foundational LLMs
24mDataEngBytes 2023 - MEL-T2-03 - Joshua Yu
30mHow to Build the Next Wave of Graph-Based Agentic AI
43mSupercharging your transformers with synthetic query generation and lexical search
49mBen Hall & Barbara Fusinska
01h 17mGraph Database and Retrieval Augmented Generation
25mTraining Large Language Models in the Open