Open-Source K-12 Accountability Analytics
A two-phase EDM/LA research plan built on a real, working K-12 education data stack. Looking for a co-author or case-study partner.
EDM 2026 accepted 126 papers. None of them use a real, open, K-12-scale dataset.
Most published EDM 2026 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.
"The single most-cited gap across the 32 full papers."
It's not a research prototype. It's a shipped system.
Open-source on GitHub. Open-weights on HuggingFace. Live demo running on HF Spaces. Real users, real failure modes, real data.
Aeries is the dominant Student Information System in California (~60% of districts). CDE publishes statewide accountability data through the California School Dashboard. Both are the actual systems a CA district IT team uses — not a synthetic stand-in.
Five storage layers — core (raw Aeries), staging (CDE-loaded), obt (one-big-table for accountability), analytics (derived marts), cross (multi-source joins). Student IDs are SHA-256 hashed. The schema is open and the framework is MIT-licensed.
Fine-tuned on hand-curated + auto-generated NL→SQL pairs over the warehouse schema. Published on HuggingFace under the KDDSTLC org in two variants (v1 r=32, v2 r=64) plus a GGUF quantization for local inference. The model is administrator-facing — the user is a school principal or analyst, not a student.
Built in Rill, the dashboards surface the four California School Dashboard indicators: Chronic Absenteeism, Suspension Rate, ELA (English Language Arts), and ELPI (English Learner Progress). Each is a drill-down view, filterable by school, reporting category, and year.
Two papers. Four months. Co-author first, benchmark second.
A case study grounds the benchmark with a real user story; the benchmark characterizes where the LLM components fail on accountability-metric semantics. Together they fill the gap EDM 2026 didn't.
Co-author search + case-study prep
Outreach to HF community, CA CS-ed faculty, and CA district IT/data folks. Reading EDM 2026 demo/industry papers for structure. Recording a 3-minute screencast of the Gradio app in action.
Case-study paper
8-page EDM format. 3 research questions: adoption, failure modes, equity. Mixed methods: 30-day log analysis + 3-5 semi-structured interviews. Four narrative arcs developed.
K12-Bench: accountability-metric reasoning
220 frozen prompts across 4 indicators × 5 difficulty tiers. 6 models evaluated (LFED LoRA, Qwen base + 32B, GPT-4o-mini, Claude Haiku, DeepSeek, Llama-70B). Failure-mode auto-tagging.
What we're actually studying
Open-source K-12 education analytics — combining a privacy-preserving DuckDB warehouse, a fine-tuned NL-to-SQL model, and state-accountability-aligned Rill dashboards — can be deployed in a district and meaningfully shift who has access to accountability data.
Three research questions
Adoption
How does a district administrator actually use the LFED stack to interrogate accountability data? What workflows does the NL→SQL layer enable that weren't possible before?
Failure modes
Where does the system break? What categories of accountability-metric questions does the LoRA fail on, and how does a non-technical user work around it?
Equity
Does open-source tooling change who asks the data questions? Does it shift from "IT team runs SQL" to "principal runs a natural-language query"?
Four narrative arcs
Click each arc to read the full story. The case study will develop Arc A and Arc C in parallel (principal + superintendent, the simplest and broadest users). Arc B is held in reserve unless the audit angle resonates with a co-author. Arc D is long-tail and requires an intervention partner.
Setting: An elementary principal at a 400-student school. Her district is on the California School Dashboard red list for chronic absenteeism. She wants to see her school's chronic absenteeism rate by grade for the last 3 years, broken out by English Learner status.
Story: The principal doesn't write SQL. IT is busy. She opens the LFED Gradio app, types the question, gets a chart. She uses the chart in a staff meeting the next morning. The chart ends up in the LCAP (Local Control and Accountability Plan) draft.
Failure mode: The LoRA returns 2023-24 rows when the principal asked for 2023-2024 attendance from Aeries. CDE uses "2023-24" format; Aeries uses "2023-2024". The principal notices the chart looks "off" relative to her mental model and manually edits the SQL in the Gradio UI.
Why this is a strong arc: Concrete, relatable, ties directly to CA Dashboard accountability. The failure mode is a real domain-semantic mismatch, not a generic SQL bug.
Setting: A district data analyst preparing for the annual equity audit. She needs to verify CDE's pre-computed Status/Change/Color against raw Aeries data for chronic absenteeism.
Story: The analyst uses the Rill dashboard to identify candidate discrepancies. For each suspect, she uses the NL→SQL app to pull raw Aeries counts. She reconciles in a Python notebook.
Failure mode: The LoRA doesn't know about the is_suppressed column. It includes suppressed rows in the average, producing a different number than CDE. The analyst spots the divergence and asks for a "filter suppressed rows" prompt — a prompt the LoRA has never seen.
Why this is a strong arc: Audit/cross-validation is a high-stakes use case. Strong equity angle. Held in reserve unless the audit angle resonates with a co-author.
Setting: Superintendent presenting to the school board tomorrow night. She needs four charts: chronic absenteeism, suspension, ELA, ELPI — all broken out by reporting category (race/ethnicity, EL status, SWD, etc.).
Story: The superintendent opens the Rill dashboard, asks the LoRA to pull the OBT (one-big-table) mart for the four indicators, gets the multi-indicator query in 30 seconds. The four charts render in the same view. She exports a PDF and sends it to the board chair.
Failure mode: Reporting category code confusion. The LoRA is asked for "Black students" but the CDE reporting category code is RB (and "African American" is sometimes a separate code). The LoRA invents an alias, returns wrong rows, the chart looks plausible but is wrong.
Why this is a strong arc: High-visibility, ties to all 4 indicators equally. The reporting category bug is a perfect example of accountability-semantic failure.
Setting: A simulator that, given a help-seeking log, decides LLM-route vs. human-route based on signals (keystroke confidence, time-on-task, prior mastery).
Status: Held in reserve. Requires a real intervention partner and empirical help-seeking data the project doesn't currently have. If a CS-ed faculty member with an active intervention study joins, Arc D becomes the natural research-design conversation.
Why this is on hold: Novelty is high, but effort is 12+ weeks. Doesn't fit the 4-6 week Phase 1 case-study window.
Three ways to be involved
The minimum viable partner is one who'll go on the byline and give 1-2 hours of interview time and a quote. But there are three roles here — pick the one that fits.
Co-Author
For CS-ed or LA researchers who'd join the byline and shape the failure-mode analysis. Realistic ask: 5-10 hours over 4-6 weeks, mostly async.
- • Authorship on a published EDM-track paper
- • Access to the warehouse schema and the eval harness
- • Co-design credit on the K12-Bench benchmark (Phase 2)
- • A worked example for your own deployment or course
- • Co-presentation rights if the paper is accepted as a talk
Reviewer
For senior researchers who'd give one round of feedback before submission. Lower time commitment, still a material contribution to acceptance odds.
- • Early visibility on the failure-mode findings
- • An open invitation to co-author Phase 2 (K12-Bench)
- • A line in the paper's acknowledgments
- • A clear-eyed view of where LMs fail on K-12 ed data
District Partner
For CA district IT directors, data analysts, or principals who'd test the system in their district for 30 days and share their experience. No SQL required.
- • A working NL→SQL interface for your accountability data
- • The Rill dashboards configured for your 4 indicators
- • Continued access after the study ends (MIT-licensed)
- • A case study that documents your workflow for the field
- • Documentation for your IT team on how to operate it
What success looks like
A real, open K-12 NL→SQL system
A reproducible schema, a working LoRA, an eval harness, and a 220-prompt benchmark. Future researchers can build on it without starting from scratch.
Authorship and a working deployment
Co-authors get a published EDM paper. District partners get a working NL→SQL interface for their accountability data, with documentation their IT team can operate.
Two papers, grounded in real users
The case study gives K12-Bench a real user story. The benchmark gives the field a reproducible failure-mode analysis. Together they fill the gap EDM 2026 didn't.
Want to talk?
If you're at a California district that might test this, even informally — reply.
If you're a CS-ed or LA researcher who'd review or co-author — 30 min on a call would shape the paper.
If you've starred the HF Space or commented on the model — I'm casting a wide net for this.
Two public-facing research notes go behind the scenes of this project. The synthesis covers the literature, the case-study plan, and K12-Bench. The strategic-pressures doc lays out the four risks we identified and the decisions we made. Both are open to feedback.
Subject: Quick question — are you using open-source ed-data tooling in your district? Hi [name], I'm reaching out because [shared context: HF community / your paper / mutual contact / district IT lead]. I'm putting together a case study paper for EDM 2027 on open-source K-12 accountability analytics. The artifact is a local-first warehouse + fine-tuned NL→SQL model + Rill dashboards, all open-source: - github.com/flucido/local-data-stack - huggingface.co/KDDSTLC/lfed-qwen2.5-coder-14b-sql-lora-warehouse-r64 - Live demo: huggingface.co/spaces/KDDSTLC/LFED I'm looking for a co-author (or just a 1-hour interview) from a CA district who's actually tried the system — IT director, data analyst, or principal who uses the NL→SQL interface to ask accountability questions. If that's you (or someone on your team), I'd love to set up a 30-min call. No prep needed — I'm just trying to understand real workflows. Best, [Your name]