Memgraph Fine-Tuned a 1B Model for Text2Cypher with $26 of Synthetic Data: 94% Synthetic, 41% Public Benchmark
JeremyCMorgan · x · 2026-11-13
A Memgraph engineer fine-tuned a 1B open-source model for Text2Cypher (translating plain-English questions into Cypher graph queries). The dataset was built backwards: Cypher queries were generated programmatically from the grammar, then DeepSeek produced varied natural-language phrasings. The final version used hundreds of synthetic schemas, 69,500 verified queries, and 270K training rows at roughly $26 in API costs; training ran overnight on a laptop GPU, and inference runs on CPU, generating some queries in under a second. Failed queries and error messages were later added so the model learns to self-correct.
The key contrast: the model hit 94% on unseen synthetic domains but only 41% on a public benchmark—a clear case of synthetic training gains failing to transfer.
More from coding & agent
- Demo: Atlas + Fable 5.2 + three.js for 3D Scene Generation — XRarchitect · 2026-11-13
- DocFS turns documents into file systems so VLM agents can reason over them — _reachsumit · 2026-11-13
- Indie devs are ditching Intercom: "you don't use 90% of the product" — jonathan_wilke · 2026-11-13
- Pinokio dev ships Zonos3-Mini Voice Studio with realtime voice cloning on Mac — cocktailpeanut · 2026-11-13
- Pinokio apps now reachable from any agent: ask Claude to clone a voice — cocktailpeanut · 2026-11-13
- Gemini 4 Hydrogen lands in Cursor, scoring 54.3% on CursorBench vs 39.6% for 3.8 Flash — mattyp · 2026-11-13