soyrsoyr/jev-playground-rlcd-v0

openjev-rlcd-v0

Synthetic calibrated-decision dataset for training Jev-style System One classifiers plan). States are generated programmatically per domain; typed questions follow the openjev /v1/systemone contract; reference distributions are exact by construction, the property a proper-scoring-rule / calibration objective needs.

Format

JSONL, one state per line:

{
 "domain": "support_tickets",
 "state": "...",
 "questions": {
  "urgent": {"type": "noul", "instructions": "...", "reference": 0.913},
  "owner": {"type": "choice", "instructions": "...", "criteria": {"billing": "...", ...}, "reference": {"billing": 0.9, ...}},
  "severity": {"type": "score", "instructions": "...", "criteria": ["level 0", "level 1", "level 2"], "reference": [0.05, 0.9, 0.05]}
 }
}

• noul → scalar P(true)
• choice → distribution over the criteria label set (sums to 1)
• score → distribution over ordered level indices 0..n-1 (sums to 1)
• reference is the ground-truth distribution, emitted by the generator, not a model label.

Domains

Domain	States	Questions
support_tickets	2500	urgent (noul), angry (noul), refund (noul), owner (choice, 4 teams), severity (score, 3 levels)
product_reviews	2500	recommends (noul), topic (choice, 4 aspects), positivity (score, 3 levels)

Uncertainty structure

Signal strength is sampled per state: crisp (70%, reference mass 0.88-0.98 on the true label), weak (20%, 0.55-0.75), contradictory (10%, 0.35-0.5 with distractor mass). This gives calibration training and ECE evaluation a full range of genuine uncertainty, not just one-hot targets.

Splits

• train.jsonl: 4500 states (40464 pairs)
• val.jsonl: 500 states (4536 pairs)

Generation

Deterministic under seed 42 via training/data.py in the openjev repo (python -m training.data --n 2500 --seed 42). The code/ directory snapshots the exact generator + trainer used.