Encoder–Decoder Project 04

Data-to-Text Executive Report Generator

Upload a synthetic CSV or JSON KPI table, compute deterministic movements, generate a structured row-grounded claim plan, verify metric, period, direction, values, ranking, risk, anomaly, and causality, then render only supported executive statements.

Input

Upload a KPI table

CSV / JSON

Maximum demo size: 1 MB, 500 rows, and 40 columns. Uploaded content is processed for the current request and is not intentionally persisted by this application.

Safe demo data

Select a sample

Explainable rules

Risk and anomaly configuration

User-controlled

Target risk is evaluated only when a metric direction is explicit in the table or configured here. This prevents the app from assuming that higher or lower values are universally better.

Visible model contribution

Generation mode

Grounding view

Uploaded table preview

6 rows
#metriccategoryperiodcurrent_valueprevious_valuetargetstatusrisk_level
1Product failure casesQualityJune 2026146118100Above targethigh
2First pass yieldManufacturingJune 202694.296.197Below targethigh
3Customer complaintsServiceJune 2026736860Above targetmedium
4Supplier defectsSupplier QualityJune 2026314635Improvinglow
5On-time closure rateQualityJune 202691.48895Below targetmedium
6Scrap rateManufacturingJune 20262.82.11.8Above targethigh

Run analysis to view KPI movements.

Structured generation

Verified executive report

Choose a generation mode, review the deterministic analysis, and generate a claim-gated report.

Traceability

Source-cell references

Select a report statement to highlight its supporting table cells.

Claim gate

Claim-level verifier

Verification results appear after report generation.

Safety control

Unsupported claims

0 blocked

Blocked model statements will appear here.

Reusable output

Export report

Model evidence

Evaluation metrics

No invented scores

The repository includes scripts for the metrics below. Values remain intentionally blank until the pretrained baseline, fine-tuned model, verifier-gated model, and optional ONNX model are evaluated on the same held-out split.

PARENTRun evaluation
BLEURun evaluation
ROUGE-LRun evaluation
BERTScoreRun evaluation
Content-selection F1Run evaluation
Numerical factualityRun evaluation
Hallucination rateRun evaluation
CoverageRun evaluation

Architecture and limits

Model and deployment details

Human review required
Primary training modelFLAN-T5-large full BF16 fine-tuning on RTX 5090
Comparison modelFLAN-T5-large LoRA plus the untouched pretrained baseline
Model outputSchema-versioned JSON claim plans with direct row IDs
Claim gateMetric, period, direction, values, ranking, risk, anomaly, causality, and sentence-number checks
Vercel inferenceSecure hosted champion endpoint with deterministic fallback
OptimizationValidated Optimum ONNX export, INT8 comparison, and latency benchmark workflow

Verification reduces unsupported claims but does not establish business causality or replace domain review. The public demo must use synthetic or public data only.