Jev Curate
Sifts synthetic and pretraining datasets with typed Jev decisions.
Added to Jevfast
How it uses Jev
The dataset pipeline streams Parquet or JSONL records and evaluates a preset rubric with Jev per row. The reasoning-math preset asks a Noul circularity question and a 0–4 rigor Score; host thresholds decide which records to retain.
What Jev decides
Reasoning-math presetExample answers · not a recorded Jev response · Source ↗
Question 1 · has circular reasoning
YOUR APP
INSTRUCTION
Does this solution contain circular reasoning or assume its conclusion?
STATE
One synthetic reasoning or instruction-tuning row from the input dataset.
JEV · NOUL
YesNo
Question 2 · reasoning depth
YOUR APP
INSTRUCTION
Rate mathematical and logical rigor.
STATE
One synthetic reasoning or instruction-tuning row from the input dataset.
JEV · SCORE
0
App workflow
- Stream input records
Read Parquet or JSONL rows and apply host-side sanity filtering.
- Score each row
Send the selected preset questions in one Jev request for each retained row.
- Separate output
Apply preset Noul and Score thresholds, writing clean and rejected records without rewriting their text.
Why it is interesting
A large curation job can apply typed, parallel rubric questions to each record without rewriting source examples, while preserving a transparent threshold for rejection.