jev-dataops
An open-source JEV-powered workbench for streaming data selection, quality evaluation, automatic LoRA training and held-out model evaluation.
Added to Jevfast
How it uses Jev
In live mode, local format, length, and duplicate checks are combined with JEV screening of dataset quality. The pipeline partitions records into keep, review, and reject, then trains the configured target model on retained data and evaluates it against a held-out set; the public demo uses local rules and does not call JEV.
What Jev decides
Training record screening requestExample answers · not a recorded Jev response · Source ↗
Question 1 · quality
YOUR APP
INSTRUCTION
Assess the clarity, completeness, and coherence of the training record.
STATE
The selected rubric receives a candidate training record as state. This chart uses the checked-in general rubric, whose three question keys and options are fixed.
JEV · CHOICE
- good
- uncertain
- bad
App workflow
- Load and preview data
Upload JSONL or CSV and record dataset details before processing.
- Screen and partition records
Run local validation and deduplication, then in live mode use JEV screening to assign keep, review, or reject.
- Evaluate the dataset
Summarize screening decisions, duplicates, and errors with a per-record audit trail.
- Train and evaluate the target model
Split retained records by groups, train the configured model, and compare loss and perplexity on the same held-out test set.
Why it is interesting
It keeps data screening and model evaluation distinct and provides per-record audit trails, making both training-data selection and before/after loss and perplexity reviewable. The README explicitly says training updates the configured model, not JEV.