The 60-second read

An open-source K-12 ed data stack — and a research plan built on it

We're building an open-source K-12 education data stack: a privacy-preserving warehouse over Aeries (the dominant Student Information System in California) and CDE (California Department of Education) public data, a fine-tuned language model for natural-language SQL queries, and Rill dashboards aligned to the four California School Dashboard indicators.

On top of the stack, we're publishing two academic papers: a case study of what happens when a district can ask accountability questions in plain English, and a benchmark that characterizes where current language models fail on K-12 education data. The case study is the first paper; the benchmark is the second.

The stack is real and live. The papers are in progress. The research is at a stage where collaborator input would change the work — we're sharing it publicly to invite that input.

Five things that are running today

• DuckDB warehouse — 30 tables, PII-hashed (student_id_hash), 5 storage layers (core / staging / obt / analytics / cross). Open schema, MIT-licensed framework.
• Fine-tuned Qwen2.5-Coder-14B LoRA for NL→SQL over the Aeries schema. Two public versions on HuggingFace (v1 r=32, v2 r=64) plus a GGUF quantization for local inference.
• Rill dashboards — 4 indicator views (Chronic Absenteeism, Suspension, ELA, ELPI), each drill-down by school, reporting category, and year. Aligned to the California School Dashboard.
• Live demo on HuggingFace Spaces — a Gradio app that lets a non-technical user ask accountability questions in plain English and get back a chart.
• Eval harness — single-model runner with execution match, exact match, JSONL traces, latency. Plus 20 hand-curated NL→SQL prompts as the dev/calibration split.

Two papers, four months, three phases

The work is sequenced so each phase produces an artifact the next builds on.

Phase 0 — co-author search + case-study prep (Jul 10 – Aug 6, 2026). The current phase. Outreach to HF community, CA CS-ed faculty, and CA district IT/data folks. Reading EDM 2026 demo/industry papers for structure. Building this public site.

Phase 1 — case-study paper (Aug 7 – Sep 10, 2026, ~5 weeks). 8-page EDM format. Mixed methods: 30-day log analysis of the NL→SQL interface, plus 3–5 semi-structured interviews with district administrators. Three research questions: adoption, failure modes, equity of access. Four narrative arcs developed as candidate case-study stories. Target venue: EDM 2027 (Industry or Demo track), with IJIET as a rolling-submission fallback.

Phase 2 — K12-Bench: accountability-metric reasoning (Sep 11 – Nov 5, 2026, ~8 weeks). A frozen 220-prompt benchmark for K-12 NL→SQL, covering the four CA Dashboard indicators across five difficulty tiers. Six language models evaluated, including the 14B LoRA. Failure-mode analysis on the domain-specific semantic errors (Status/Change/Color, suppression, reporting category codes, year-format mismatches). Target venue: EDM 2027 Full Paper.

The gap in the EDM/LA literature

EDM 2026 accepted 126 papers. None of them use a real, open, K-12-scale dataset. Most published work draws on higher-ed data (OULAD, ASSISTments, Khan Academy) or controlled lab studies. The single most-cited gap across the 32 full papers: we have no shared, open, real-world K-12 systems that practitioners can build on.

The few K-12 datasets that exist are proprietary. The few K-12 systems that exist are vendor-locked. There is no open K-12 NL→SQL stack for the community to study. That's the gap this project fills.